TL;DR

• New speech models can handle live multilingual conversation, including continuous Chinese–English translation and real-time speech apps with web search or function calling.
• Several new models focus on reading documents, from extracting receipt totals to parsing multi-page files and tables.
• Liquid AI released models for on-device voice and video tasks, as well as local agents and media classification.
• Maya 2 Native offers streaming speech generation in 12 languages, with open weights and voice cloning across languages.

New models bring live speech translation and voice features

Hy-SimulMT-30B-A3B and Hy-SimulMT-30B-A3B-FP8 translate incoming Chinese or English speech into text and spoken translation as a conversation unfolds. Their voice-cloning feature can preserve the speaker’s vocal characteristics, but it requires a configured text-to-speech service. Qwen-Audio-3.1-Realtime-Flash is built for two-way speech interactions: it can take in and produce both text and audio, and supports web search and function calling. Developers could use those tools to build speech apps that look up information or trigger actions during a conversation.

Maya 2 Native generates streaming speech in 12 languages and includes 20 voices. Its open weights and inference code can support locally hosted voice applications, while cross-lingual voice cloning lets it generate speech in another language using a voice sample.

New OCR models target receipts, tables and long documents

MiniCPM-OCR 1.0 is designed to read more than plain text: it can handle tables, formulas, charts and other document elements in Chinese and English. One practical use is extracting a specific field, such as the total on a receipt. Unlimited OCR 2 is aimed at parsing images and long, multi-page documents, including PDFs. Its description says it can process documents beyond a finite context window, but does not give a specific page limit.

Liquid AI adds models for local media tasks and agents

LFM3-Omni-2B can take in text, images, video and audio, then return text, speech or both. The model is designed for on-device use; listed applications include transcribing or translating speech and asking questions about a video’s visuals or soundtrack. d1-Omni-2B is a decision model that can classify or extract information from text and media, such as locating objects in images or routing calls based on audio. Instead of generating a written response, it returns answers with probabilities. LFM3-1.7B is an open-weight text model intended for local agents and specialized applications. LFM3-1.7B-Base is a separate pre-trained checkpoint intended as a starting point for substantial fine-tuning.

Also on the radar

• Ling-3.1-flash-base is an open-weight base model for continued pretraining, adapting a model to a domain, and other research. It is intended as a starting checkpoint rather than a finished assistant.
• Chaos Intelligence is a programming language and development environment for building programs that combine functions, language models and agent workflows.

The records

• Ling-3.1-flash-base (inclusionAI): Ling-3.1-flash-base is a base checkpoint released to support research and community-driven innovation. It has 560B total parameters and 25B active parameters per token, with a hybrid-linear Mixture-of-Experts architecture.
• Maya 2 Native (Maya Research): Maya 2 Native is a streaming text-to-speech model that generates 24 kHz speech in 12 languages, with 20 built-in voices. It includes open weights and inference code, supports zero-shot cross-lingual voice cloning, and is released under Apache-2.0.
• Qwen-Audio-3.1-Realtime-Flash (Alibaba Qwen): Qwen-Audio-3.1-Realtime-Flash is a real-time duplex speech model that accepts and produces text and audio. It supports function calling, web search, voice cloning, and a 262,144-token context window.
• LFM3-Omni-2B (Liquid AI): LFM3-Omni-2B is a post-trained omni-modal model that accepts text, images, video, and audio, and produces text, speech, or interleaved text and speech. The 2.17B-parameter model is designed for on-device use and has downloadable BF16 and GGUF checkpoints.
• d1-Omni-2B (Liquid AI): d1-Omni-2B is a 2.17B-parameter decision model built on LFM3-Omni-2B that answers typed questions in one forward pass without generating tokens. It accepts text, images, video frames, and audio, and returns answers with probabilities.
• LFM3-1.7B (Liquid AI): LFM3-1.7B is a text model in Liquid AI’s LFM3 family, built for local agents and specialized applications. It has a 256K-token context window, is post-trained for agentic harnesses, and is available as open weights.
• LFM3-1.7B-Base (Liquid AI): LFM3-1.7B-Base is a pre-trained, text-only 1.73B-parameter checkpoint in Liquid AI’s LFM3 family. It has a 256K-token context length and is intended as a starting point for substantial fine-tuning.
• MiniCPM-OCR 1.0 (OpenBMB): MiniCPM-OCR 1.0 is a 1.3B OCR-focused model for understanding documents, including text, tables, formulas, charts, seals, and structured information. It supports Chinese and English and is released with model weights, API documentation, SDK support, and deployment guides for several runtimes.
• Hy-SimulMT-30B-A3B (Tencent Hunyuan): Hy-SimulMT-30B-A3B is a real-time simultaneous translation model for bidirectional Chinese–English speech translation. It continuously processes incoming audio and produces translated text and speech, with real-time voice cloning supported when a TTS service is configured.
• Hy-SimulMT-30B-A3B-FP8 (Tencent Hunyuan): Hy-SimulMT-30B-A3B-FP8 is a real-time simultaneous translation model for bidirectional speech-to-speech translation between Chinese and English. It continuously processes incoming audio, produces translated text and speech, and supports real-time voice cloning; speech output and voice cloning require a configured TTS service.
• Unlimited OCR 2 (Baidu ERNIE): Unlimited OCR 2 is an image-text model for parsing documents, including multi-page documents. It uses Sparse Reference Sliding Window Attention to support unlimited-horizon parsing within a finite context window.
• Chaos Intelligence (Cattmonkey): Chaos Intelligence is a functional programming language that treats AI agents as callable values, with a CLI and web IDE. It supports building agent workflows, goal-driven bots, and interactive pages and reports.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• Lure 0.6B can create expressive English speech from a short voice sample and transcript, with optional directions such as how to deliver a line.
• d3-ultra can make decisions about text, images, and video, including answering yes-or-no questions, scoring, and locating or tracking objects.
• App Studio turns plain-language descriptions, brand materials, and files into hosted creative apps.
• Carbon-X is an open model for researchers working with DNA sequences and natural-language prompts.

Create expressive speech from a short voice sample

Lure 0.6B turns text into English speech and can clone a voice using a 3–12 second reference clip and its transcript. It supports streaming audio, sound tags such as sighs and laughter, and optional directions for how the speech should sound. That could help people making voice content create a performance in a supplied voice without recording every line. The model is available through the web, an API, and open weights.

A model that can classify, score, or track what it sees

d3-ultra accepts text or structured questions, optionally with images or video, and returns answers with probabilities rather than writing a free-form response. It can choose among options, answer yes-or-no questions, score against criteria, locate objects in images, and track them in video. This is aimed at developers building apps that need an AI to make a specific decision—for example, identifying an object or routing an input—rather than hold a conversation. It is available on the web and through an API, with open weights.

Build and host a creative app from a description

App Studio builds creative apps from plain-language descriptions, using Runway’s models along with a user’s brand materials and files. It also hosts the apps it creates. It may suit people who want a custom creative tool without building and hosting one from scratch.

An open model for working with DNA and language

Carbon-X is an open model trained on DNA sequences and natural-language text. Researchers can prompt it with DNA and ask it to respond in English, DNA, or both, for tasks such as examining sequence features, predicting the effects of variants, or predicting the next DNA base. It is intended for researchers who want to run DNA-analysis experiments on their own infrastructure.

Also on the radar

• LBS-27B inspects image regions adaptively and is intended for visual question answering and locating interface elements; a client must carry out its image-cropping actions.
• Nemotron-3.5-Embed-8B-BF16 converts text queries and text or image documents into vectors for multilingual and multimodal search and retrieval.

The records

• Lure 0.6B (Tesslate): Lure 0.6B is an English text-to-speech model that can clone a voice from a 3–12 second reference clip and its transcript. It supports streaming audio, inline sound tags, and optional free-text acting directions.
• d3-ultra (vLLM Semantic Router): d3-ultra is a 27B multimodal decision model that accepts text or JSON, optionally with images and videos, and returns answers with probabilities rather than generating text. It supports choice, yes/no, scoring, box localization, and video tracking decisions, with up to 131,072 tokens per question.
• LBS-27B (inclusionAI): LBS-27B is an agentic vision-language model based on Qwen3.8-27b. It uses adaptive image-region inspection learned from synthetic GUI-grounding data and is released as a checkpoint.
• Nemotron-3.5-Embed-8B-BF16 (NVIDIA): Nemotron-3.5-Embed-8B-BF16 is an 8.38B-parameter embedding model for multilingual and multimodal retrieval. It accepts text queries and text, image, or combined image-and-text documents, and produces embedding vectors.
• Carbon-X (Hugging Face): Carbon-X is an open 14.8B-parameter decoder-only language model trained jointly on DNA sequences and natural-language text. It accepts DNA in chat prompts and can respond in English, DNA, or both, with capabilities including variant-effect prediction, sequence understanding, and next-base prediction.
• North-Mini-Code-1.0 (Cohere): North Mini Code (North-Mini-Code-1.0) is an open-source mixture-of-experts agentic coding model from Cohere that uses 30B total parameters with 3B active to deliver efficient software-engineering capabilities. It is released under the Apache 2.0 license with downloadable weights and multiple hosted and API access options.
• RFD4-Proteína (NVIDIA): RFD4-Proteína is a flow-matching model that generates protein sequences and all-atom structures together. It can condition designs on molecular context such as a binding target, enzyme motif, ligand, or selected structural features.
• North-Small-Translate-1.0 (Cohere): North-Small-Translate-1.0 is an open-weight mixture-of-experts machine translation model from Cohere optimized for high-throughput, long-context translation across 50+ languages. Cohere publishes the weights on Hugging Face under a CC BY-NC 4.0 license and provides a demo space and implementation guides for research and non-commercial use.
• GPT-6.1 Terra (OpenAI): GPT-6.1 Terra is described as a fast, efficient model for visual reasoning, precise computer use, and clear, natural writing. It is available through the Responses and Chat Completions APIs.
• North 2 (Cohere): North 2 is a full-stack agentic platform for enterprises, with a redesigned orchestration system for multi-step agents and automations. It includes reusable skills, shared libraries, agent memory, app creation, connectors, governance controls, and flexible deployment options.
• Claude Platform (Anthropic): Anthropic updated the Claude Platform to let the server automatically evaluate Managed Agents tool calls via a new "auto" permission mode and to report evaluation metadata in agent events. The ant CLI (v1.32.0) gained a beta sessions connect command to attach a terminal to Managed Agents sessions and a --web option to open the session viewer locally.
• App Studio (Runway): App Studio builds creative apps from plain-language descriptions using Runway’s models, a user’s brand and files. It hosts the apps it creates.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• Granite Speech 5.0 2B can transcribe multilingual speech and translate speech to and from English, with language prompting and larger keyword lists.
• OlmoEarth v1.3 Base turns satellite time series into map-ready features, and the OlmoEarth Embeddings Explorer lets people inspect and classify those features in a browser.
• Hy-Nomas generates video in successive segments that can be extended with text instructions and camera control.
• WeVisOne-VL-4B can locate and count objects in images, as well as handle tasks such as segmentation and image matting.
• Several new open-weight models target narrower jobs, including healthcare research, shorter reasoning, and rewriting Chinese web novels.

A speech model adds multilingual transcription and translation

Granite Speech 5.0 2B is a compact speech-language model for transcribing speech and translating speech to and from English across supported languages. It adds language-identification prompting and support for larger keyword lists, which can help guide transcription toward relevant terms; IBM also reports improved transcription accuracy and robustness over Granite Speech 4.1 2B.

Explore satellite changes in a browser, or use the data to build maps

OlmoEarth v1.3 Base reads yearly radar and multispectral satellite data and produces a set of numerical features for each 10-meter pixel on Earth. Those features can support maps and analysis of land cover, crops, fires, and other landscape features when paired with small classifiers. The browser-based OlmoEarth Embeddings Explorer provides a way to inspect maps, compare years, find similar places, train a small classifier from labeled points, and export results without setting up a local environment or GPU.

Hy-Nomas lets you extend a generated video world one segment at a time

Hy-Nomas generates video in segments and uses cached context to continue the same world across turns. People can use text instructions and camera control to request changes, actions, styles, or new views in successive segments.

WeVisOne-VL-4B handles image tasks from object counting to matting

WeVisOne-VL-4B can locate, count, read, segment, or extract objects in images. It also supports image matting, which separates a subject from its background to produce an alpha matte.

Three open-weight models focus on healthcare, shorter reasoning, and Chinese fiction

Ling-3.0-flash-Sante is designed for healthcare question answering, research, evidence synthesis, and long-document analysis, with support for tool use. Its listed uses are research-oriented, and the source calls for appropriate human review. ThinkingCap Gemma-4-26B-A4B-it is a post-trained model intended to use fewer reasoning tokens; its model card reports a 38.9% average reduction and 75.0% average accuracy, compared with 75.2% for the base model. novel-humanize-4b rewrites Chinese web-novel chapters in a more human-like style, with instructions to preserve plot, dialogue, characters, and event order while changing the prose.

Also on the radar

• Groot-19B-A3B-261023 is an open-weight multimodal reasoning model with tool calling and a 262,144-token context; it is intended for reasoning, coding, and tool-using tasks, including local deployment on supported machines.
• The local inference engine rec offers an API compatible with OpenAI and Anthropic chat-completion formats, while OPAL lets a coding agent rewrite optimizer and analysis programs during a PyTorch training run.

The records

• Granite Speech 5.0 2B (IBM): Granite Speech 5.0 2B is a compact speech-language model for multilingual automatic speech recognition and bidirectional automatic speech translation. Compared with Granite Speech 4.1 2B, it improves transcription accuracy and robustness, adds language-ID prompting, and supports larger keyword lists.
• Groot-19B-A3B-261023 (GrootTeam): Groot-19B-A3B-261023 is an open-weight, natively multimodal mixture-of-experts reasoning model with about 19 billion total parameters and 3 billion activated per token. It supports a native 262,144-token context, tool calling, and BF16, FP8, NVFP4, and GGUF weight editions.
• ThinkingCap Gemma-4-26B-A4B-it (BottleCap AI): ThinkingCap Gemma-4-26B-A4B-it is a post-trained version of Google’s Gemma-4-26B-A4B-it, designed for shorter reasoning. The model card reports an average 38.9% reduction in reasoning tokens and average accuracy of 75.0%, compared with 75.2% for the base model.
• novel-humanize-4b (gugujiji): novel-humanize-4b is a Qwen3-4B-based model for rewriting Chinese web-novel chapters in a more human-like writing style. Its instructions aim to preserve the plot, dialogue, characters, and event order while changing the prose.
• OlmoEarth v1.3 Base (Ai2): OlmoEarth v1.3 Base is a satellite embedding model that reads annual radar and multispectral satellite time series. It produces a 128-dimensional vector for every 10-meter pixel on Earth, with the first 64 dimensions usable as a smaller embedding.
• Hy-Nomas (Tencent Hunyuan): Hy-Nomas is a video world generation model that supports multi-turn interaction through text instructions and camera control. It generates video a segment at a time, using cached context to extend the same world.
• Ling-3.0-flash-Sante (inclusionAI): Ling-3.0-flash-Sante is a healthcare-specialized model based on Ling-3.0-flash, with 124B total parameters and 5.1B active parameters per token. It supports long-context reasoning, tool use, and research workflows, and its weights are available on Hugging Face and ModelScope.
• WeVisOne-VL-4B (Tencent Hunyuan): WeVisOne-VL-4B is a vision-language model developed by WeChat Vision, Tencent, built on Qwen3-VL. It supports object-level visual perception and pixel-level tasks including segmentation and image matting.
• rec (Recursive): rec is Recursive’s custom inference engine for running its language models locally on a machine. It provides a local API compatible with OpenAI and Anthropic chat-completion formats and supports single-session usage.
• OPAL (Xingjian Bai): OPAL is a PyTorch optimizer wrapper that lets a coding agent rewrite optimizer and analysis programs while a model trains. It runs in the training process and supports agent-free dry runs with the `keep` backend.
• neuron explorer (lyramakesmusic): neuron explorer is a steering-first neuron explorer for Qwen3.5-4B that visualizes model layers and MLP neurons in a 3D view. It supports live steered chat, neuron activation analysis, lenses, and connections using a locally loaded model.
• OlmoEarth Embeddings Explorer (Ai2): OlmoEarth Embeddings Explorer is a browser-based interactive analysis tool built from OlmoEarth embeddings. It lets users inspect maps, compare years, find similar places, train a small classifier from labeled points, and export results as a GeoTIFF.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• AWS Continuum can trace code vulnerabilities through an AWS environment, test findings in an isolated sandbox, and propose fixes as pull requests.
• GitHub’s app for Microsoft Teams lets people ask GitHub to investigate an issue or prepare a fix from a chat.
• Cohere’s North 2 adds tools for creating, scheduling, testing, and governing multi-step automations.
• Two new language models target different work: Recursive Flint focuses on coding tasks, while NVIDIA’s Nemotron 3.5 Bolt E4B 4GB is designed for low-memory assistants.

AWS Continuum can investigate code vulnerabilities and propose fixes

AWS Continuum for code vulnerabilities looks for security weaknesses in code, checks findings in an isolated sandbox, and ranks them using information about the AWS environment. It can propose remediation patches as pull requests, or apply them automatically within configured guardrails. This may help teams focus on vulnerabilities that matter in their environment and review proposed fixes. The record describes its intended capabilities, not results from independent testing.

GitHub issues can lead to proposed fixes inside Microsoft Teams

The GitHub app in Microsoft Teams brings GitHub notifications and Copilot actions into chats. Users can mention GitHub in response to an issue notification to ask it to investigate the issue, fix a bug, or create a pull request. That could help teams move from discussing an issue to preparing a code change without leaving the conversation; the proposed work still needs appropriate review.

North 2 adds ways to build and manage multi-step automations

Cohere’s North 2 adds workflow tools for creating automations in plain language, scheduling runs, and setting up branches, approvals, and reviews. Teams can also test and version workflows, choose models, and monitor usage. The platform is aimed at organizations building agents that use internal knowledge and tools, with governance and deployment options for different security needs.

Two new models target coding work and low-memory assistants

Recursive Flint is an open-weights language model for text and code that supports multi-step reasoning and tool use. Its stated uses include investigating codebases, fixing bugs, implementing features, and writing tests. Nemotron 3.5 Bolt E4B 4GB, from NVIDIA, is designed to run with low memory and low time to first response. Its intended uses include local voice or device assistants, interactive personas, and embedded interfaces; the record does not establish performance in those settings.

Also on the radar

• Hugging Face’s torch-neuronx model support Space checks whether popular Hub models run unmodified on AWS Trainium and provides test results and instructions for supported models.
• GitHub added per-user settings and more detailed controls for automatic Copilot code reviews; its JetBrains plugin also adds agent tools, including public-preview approvals for assisted tool use.

The records

• Nemotron 3.5 Bolt E4B 4GB (NVIDIA): Nemotron 3.5 Bolt E4B 4GB is a quantized text-generation model post-trained by NVIDIA from Gemma-4-E4B-it, with 8B total parameters and 4B effective parameters. It is designed for low-memory, low-time-to-first-token operation and emphasizes multi-turn tool use, instruction following, and input grounding.
• Recursive Flint (Recursive): Recursive Flint is a mixture-of-experts language model with 39.8 billion total parameters and approximately 2.5 billion active parameters per token. It supports text input and output, reasoning, and multi-turn tool use.
• North 2 (Cohere): North 2 is a full-stack agentic platform for enterprises, with a redesigned orchestration system for multi-step agents and automations. It includes reusable skills, shared libraries, agent memory, app creation, connectors, governance controls, and flexible deployment options.
• SageMaker AI endpoints (AWS): SageMaker AI endpoints are AWS-managed inference endpoints for deploying generative AI models where AWS handles GPU provisioning, autoscaling, and operational monitoring. In 2026 AWS introduced multiple enhancements including automated inference recommendations, multi-instance capacity-aware pools, an OpenAI-compatible API surface, container caching to reduce cold-starts, token-level observability, inline async payloads, and prefix-aware routing for KV cache reuse.
• torch-neuronx model support Space (Hugging Face): The Space lets users check whether popular Hugging Face Hub models run unmodified on AWS Trainium and provides a script to run supported models. It reports test results, Trainium core counts, run times, peak device memory, and the tested software stack.
• GitHub Copilot (GitHub): GitHub deprecated four models across its Copilot experiences on October 2, 2026. It suggests replacement models and says administrators may need to enable access to alternatives through model policies.
• GitHub app in Microsoft Teams (GitHub): The GitHub app in Microsoft Teams combines GitHub notifications with Copilot actions in channels, group chats, and private chats. Users can mention GitHub to act on an issue notification, including investigating an issue, fixing a bug, and creating a pull request.
• AWS Continuum for code vulnerabilities (AWS): AWS Continuum for code vulnerabilities discovers, validates, prioritizes, and remediates code vulnerabilities. It builds a context graph from an AWS environment, tests findings in an isolated sandbox, and can provide remediation patches as pull requests.
• Copilot code review (GitHub): Copilot code review adds a dedicated per-user settings page and more granular automatic-review controls, plus an enterprise-level default review effort that can be inherited by organizations. These changes are generally available across Copilot plans including Business and Enterprise.
• Copilot for JetBrains 1.18.0 (GitHub): Copilot for JetBrains 1.18.0 is a JetBrains IDE plugin update that adds new agent features, organization-shared skills and instructions, and workflow improvements. The release includes public-preview assisted tool approvals, a re-edit message capability, Codex agent plan mode, and controls for MCP tools and the built-in MCP server.
• Copilot CLI (GitHub): Copilot CLI adds whole codebase indexing (WCI) for C++ that builds a persistent symbol index across a project so the Microsoft C++ Language Server can resolve symbols and relationships without repeatedly rediscovering them. The feature is enabled by default, can increase initial build time and memory usage, and the index is reused and dynamically updated thereafter.
• Copilot app (Local sandboxing) (GitHub): Local sandboxing for the Copilot app provides per-project policies that limit filesystem, network, and credential access for local repository and working-tree sessions. The feature is available in public preview and is configurable in the app per project or for individual sessions with an in-session command.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• Claude Models on Google Cloud lets developers access the Claude API through a Google Cloud project, with access controls and billing handled there.
• The Claude Platform adds an automatic mode for evaluating Managed Agents tool calls and new ways to connect to and monitor agent sessions.

Access Claude through a Google Cloud project

Claude Models on Google Cloud (Anthropic managed) makes the Claude API available through a Google Cloud project. Google Cloud IAM controls who can access it, while Google Cloud Marketplace handles billing. Developers can also use Fast mode with Claude Opus 5.5.

The Claude Platform can evaluate agent tool calls automatically

An update to the Claude Platform adds an “auto” permission mode that lets the server evaluate Managed Agents tool calls, and includes evaluation details in agent events. The ant CLI also adds a beta command to connect a terminal to a Managed Agents session, plus an option to open the session viewer locally.

These changes give developers and operators ways to automate or gate tool use and to monitor and approve tool calls in real time.

The records

• Claude Models on Google Cloud (Anthropic managed) (Anthropic): Claude Models on Google Cloud (Anthropic managed) provides access to the Claude API through a Google Cloud project. Google Cloud IAM controls access, and Google Cloud Marketplace handles billing.
• Claude Platform (Anthropic): Anthropic updated the Claude Platform to let the server automatically evaluate Managed Agents tool calls via a new "auto" permission mode and to report evaluation metadata in agent events. The ant CLI (v1.32.0) gained a beta sessions connect command to attach a terminal to Managed Agents sessions and a --web option to open the session viewer locally.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• NVIDIA released Flora-captioner-27B, a research model that writes an English action instruction for a selected interval in a manipulation video, to help researchers build robot datasets.

A model writes action labels for robot videos

NVIDIA’s Flora-captioner-27B generates one English instruction describing the action in a selected part of a robot or human manipulation video. It uses nearby video context to help distinguish what operation is taking place. The model is intended for research and development, and NVIDIA says it is released to help researchers scale up custom robot datasets. It is for non-commercial use, so it is most relevant to people working on robotics data collection and curation.

The records

• Flora-captioner-27B (NVIDIA): Flora-captioner-27B generates one English action instruction for a target interval in a robot or human manipulation video, using neighboring video context to disambiguate the operation. NVIDIA describes it as a research and development model for non-commercial use, released to help researchers scale up custom robot datasets.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• Fibo 2.0 Dev generates and edits images, while Fibo 2.0 Edit Turbo is designed for targeted edits guided by instructions, masks, or reference images.
• Spark-X2.6-4B is a downloadable language model for local text, coding, and tool-using tasks; its release describes improvements over the previous version.
• Ming-3.0-asr-tiny can transcribe Chinese and English speech, with prompts to help handle specialized terms or conversation context.
• AWS launched TorchNeuron in public beta for building and testing PyTorch models on Trainium hardware.

Two Fibo models offer image generation and more targeted edits

Fibo 2.0 Dev is an image model for creating images from text and editing existing images; it is also designed to be fine-tuned for specific workflows. Fibo 2.0 Edit Turbo focuses on edits, including changes guided by a written instruction, a selected area, or as many as four reference images.

Both models are available with open weights. The Dev model is the full base model; Edit Turbo is a distilled version designed to make edits in four steps.

Spark-X2.6-4B is available to run locally

Spark-X2.6-4B is a compact, general-purpose language model released as downloadable weights. Developers can run it locally for text generation, coding, and tasks that involve tools or multi-step planning. The release describes improvements in coding, tool-using workflows, longer-context tasks, multilingual abilities, and reasoning compared with Spark-X2.5-4B.

Ming model transcribes Chinese and English speech

Ming-3.0-asr-tiny is an open-weight speech recognition model for transcribing Chinese and English, including regional dialects. It accepts prompts with useful terms or preceding dialogue, which can help when transcribing speech about a specialized topic or following a conversation.

TorchNeuron brings native PyTorch support to Trainium

TorchNeuron, now in public beta, is an open-source PyTorch backend for AWS Trainium hardware. Developers and researchers can use standard PyTorch APIs to build and experiment with models on Trainium. It supports both eager execution and torch.compile, along with a way to add custom operations and optimizations.

Also on the radar

• Ornith-1.5-122B-NVFP4 is an open-weight version of a 122B-parameter model intended for text, coding, reasoning, and agent tasks; its model card describes the family as designed for single-GPU deployment.
• Rerank 5 Pro ranks documents for retrieval and can return scored passages to help identify material relevant to a query.
• EAGLE-4 draft for Qwen3.8-27B and Raon-1 mini (DFlash2 draft) are draft models intended to speed up generation when paired with their respective target models; neither is a standalone model.
• eagle-harness is an installable package for serving EAGLE-4 drafts through supported inference engines.

The records

• Fibo 2.0 Dev (Bria AI): Fibo 2.0 Dev is an 8B-parameter image model for text-to-image generation and image editing. It is the full, non-distilled base model, trained on structured captions and designed for fine-tuning.
• Spark-X2.6-4B (XHToken): Spark-X2.6-4B is a compact, general-purpose language model released as downloadable model weights in Hugging Face Transformers format. The release describes improvements in coding, agentic workflows, long-context tasks, multilingual capabilities, and reasoning compared with Spark-X2.5-4B.
• Ming-3.0-asr-tiny (inclusionAI): Ming-3.0-asr-tiny is a speech-recognition model for transcribing Chinese and English, including regional dialects. It supports context prompts with hotwords or preceding dialogue and is released with model weights, a tokenizer, custom Transformers code, and a runnable inference example.
• Fibo 2.0 Edit Turbo (Bria AI): Fibo 2.0 Edit Turbo is an 8B-parameter image-editing model distilled from Fibo 2.0 Dev. It supports instruction-based, masked, and multi-reference editing, and is designed to edit images in four steps.
• Ornith-1.5-122B-NVFP4 (Ornith AI): Ornith-1.5-122B-NVFP4 is an open-weight, NVFP4 version of Ornith-1.5-122B, a 122B mixture-of-experts model. The model card describes the family as designed for efficient single-GPU deployment and lists an MIT license.
• EAGLE-4 draft for Qwen3.8-27B (EAGLE Corp): EAGLE-4 is a speculative-decoding draft model for accelerating large language model inference. The announced draft checkpoint targets Qwen3.8-27B and is available on Hugging Face.
• Raon-1 mini (DFlash2 draft) (KRAFTON): Raon-1 mini (DFlash2 draft) is an open-weight draft model for speculative decoding with Raon-1 mini preview. It is not a standalone model and is provided to accelerate inference.
• Rerank 5 Pro (Cohere): Rerank 5 Pro is optimized for high ranking quality, especially for complex, reasoning-heavy queries and long or domain-specific documents. It supports over 100 languages, a 32k-token context window, and optional Context Focus, which returns relevant passages as scored character spans.
• TorchNeuron (AWS): TorchNeuron is an open-source native PyTorch backend for AWS Trainium, launched in public beta. It supports eager execution, torch.compile, standard PyTorch APIs, and integration with the Neuron Kernel Interface for custom operators and optimization.
• Amazon Bedrock Managed Agents (AWS): Amazon Bedrock Managed Agents is built on a customized version of OpenAI's Agents API and integrated with AWS resources. It manages agent state, tool use, code execution, and multi-step coordination, with durable sessions, reusable skills, and support for MCP servers.
• eagle-harness (EAGLE Corp): eagle-harness is a pip-installable package for serving EAGLE-4 drafts through SGLang, vLLM and TokenSpeed engine plugins. It sizes speculative trees using measured draft and verification costs for the running batch size, hardware and workload.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• Voxtral Voice Agent SDK lets developers build voice agents that can listen, use tools, and complete tasks during live conversations.
• Coral NPU and Coral Compiler v1.0 give hardware teams an open edge-AI architecture and developers a way to compile models for it.
• AnyLanguageModel 1.0 gives Swift apps one API for choosing among system, open-weight, and hosted language models.
• Ropedia W0 generates robot-oriented video from an image, a camera path, and a skeleton video.

Voice agents can take actions during a live conversation

Voxtral Voice Agent SDK combines speech recognition, reasoning, tool use, and spoken responses in a real-time pipeline. Developers can use it to build phone or web agents that retrieve information or carry out multi-step tasks through business systems.

A new Coral toolkit targets AI that runs on edge devices

Coral NPU is an open architecture for building edge-AI hardware, while Coral Compiler v1.0 lets developers compile models from JAX, PyTorch, and TensorFlow for Coral NPU devices. The compiler is intended to make it possible to compile and benchmark models without writing custom assembly. Coralboard SL and Coralboard SR are development kits for prototyping Linux-based IoT and robotics applications, and MCU-class edge-AI applications, respectively.

One Swift API can connect an app to different language models

AnyLanguageModel 1.0 gives Swift developers a shared interface for using system, open-weight, or hosted language models. It also supports tool calling, image prompts, and Linux, which may help developers build apps that can switch between model providers without changing their whole integration.

Ropedia W0 turns camera paths and skeleton motion into video

Ropedia W0 generates video conditioned on camera movement and projected skeletons. Its released checkpoints include one trained on human video and another further trained on a mix of human and robot video; a stated use is generating robot-oriented video from an image, camera trajectory, and skeleton video.

Also on the radar

• Ming-3.0-tts-tiny is a downloadable speech model for multilingual speech, voice cloning, and control over speaking style; its stated release date is 22 September, not this brief’s date.
• TorchTitan on TPU and TorchTPU add ways to train or run PyTorch workloads on TPUs, while DeepInfer is an inference framework for serving supported DeepSeek models. Beam 1.0 501B A23B supports text, code, reasoning, and tool use, but its stated release date is 7 October.

The records

• Beam 1.0 501B A23B (Reflection): Beam 1.0 501B A23B is an instruction-tuned, open text model with a sparse mixture-of-experts architecture and 501B total parameters. It supports a 1M-token context window, reasoning, software engineering, and agentic tool use.
• Ropedia W0 (Ropedia): Ropedia W0 is an egocentric world foundation model for generating video conditioned on camera poses and projected skeletons. Its released 5B checkpoints are based on the Wan2.2-TI2V-5B backbone and were trained on human egocentric video, with one checkpoint further trained on a human-and-robot mixture.
• Ming-3.0-tts-tiny (inclusionAI): Ming-3.0-tts-tiny is a downloadable mixture-of-experts speech synthesis checkpoint built on the Ling V3 backbone, with 7.9B total parameters and about 1.3B active per token. It supports multilingual speech generation, voice cloning, voice design, dialogue, and controls for speaking style and delivery.
• Coralboard SL (Synaptics): Coralboard SL is a Linux-based IoT and robotics development board powered by the Astra SL2619 processor with a 1 TOPS Torq/Coral NPU subsystem. It is generally available worldwide through DigiKey.
• Coralboard SR (Synaptics): Coralboard SR is a developer kit powered by the Astra SR200 Series. It is designed for prototyping MCU-class Edge and Physical AI applications.
• Coral Compiler v1.0 (Google): Coral Compiler v1.0 is an open-source IREE/MLIR compiler and simulation stack for targeting Coral NPU hardware. It accepts models from JAX, PyTorch, and TensorFlow and lowers workloads to Coral NPU instructions.
• Coral NPU (Google): Coral NPU is an open, standards-based RISC-V architecture for edge AI. Its completed 32-bit design includes a Matrix Execution Unit configurable for up to 1 TOPS.
• TorchTitan on TPU (Google): TorchTitan on TPU is a Google fork of Meta’s TorchTitan adapted for training models on Google Cloud TPUs. It provides TPU-native distributed training recipes and optimization features, including full training-step graph capture and configurable Pallas kernels.
• AnyLanguageModel 1.0 (Hugging Face): AnyLanguageModel 1.0 is a Swift API compatible with Apple's Foundation Models API that lets developers choose among system, open-weight, and hosted models. It supports tool calling, guided generation, image prompts, multiple inference providers, and Linux.
• TorchTPU (Hugging Face): TorchTPU adds a native TPU backend to PyTorch. It supports Pallas kernels, distributed training, and torch.compile integration.
• Voxtral Voice Agent SDK (Mistral): The Voxtral voice agent SDK combines listening, reasoning, tool calling, and speech in a real-time streaming pipeline. It lets agents interact with business systems and complete multi-step tasks during live conversations.
• DeepInfer (DeepSeek): DeepInfer is an open-source inference framework for serving DeepSeek models at scale, with offline inference and features such as optimized kernels, asynchronous scheduling, and structured outputs. It supports DeepSeek-V4.1-Flash and runs on NVIDIA SM100 GPUs or Huawei Ascend 950 NPUs.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• OpenGCUA for macOS from Tongyi-MAI: OpenGCUA for macOS is a menu-bar assistant with a floating task panel, voice, and a bundled harness, SDK, and driver.
• solidcv from solidSF: solidcv is a pure-Rust computer vision and spatial-computing library that also provides tools for VLM token pruning and mechanistic interpretability.
• OpenGCUA Cross-Device SDK from Tongyi-MAI: OpenGCUA Cross-Device SDK provides a programming interface for computers, browsers, and phones.

Also on the radar

• CloudWatch Omni (AWS): CloudWatch Omni is an AI-powered observability experience that centralizes telemetry, dependency mapping, and troubleshooting across AWS accounts, Regions, and other clouds.
• BiRefNet v2 (Runway): BiRefNet v2 is available in Runway MCP for automatic background removal.
• Mirelo SFX 2.0 (Runway): Mirelo SFX 2.0 is available in Runway MCP on paid plans.
• Copilot code review (GitHub): Copilot code review adds a dedicated per-user settings page and more granular automatic-review controls, plus an enterprise-level default review effort that can be inherited by organizations.
• GitHub Copilot (GitHub): GitHub deprecated four models across its Copilot experiences on October 2, 2026.
• Copilot CLI (GitHub): Copilot CLI adds whole codebase indexing (WCI) for C++ that builds a persistent symbol index across a project so the Microsoft C++ Language Server can resolve symbols and relationships without repeatedly rediscovering them.
• Copilot for JetBrains 1.18.0 (GitHub): Copilot for JetBrains 1.18.0 is a JetBrains IDE plugin update that adds new agent features, organization-shared skills and instructions, and workflow improvements.
• LiteRT Tensor API (Google): LiteRT Tensor API is a C++ authoring layer for building and optimizing on-device model computation graphs.
• Qwen3.8-Flash-Next-DFlash2 (Inco AI): Qwen3.8-Flash-Next-DFlash2 is a block-diffusion draft model for speculative decoding with Qwen3.8-Flash-Next.

The records

• BiRefNet v2 (Runway): BiRefNet v2 is available in Runway MCP for automatic background removal. It detects the foreground on its own.
• Mirelo SFX 2.0 (Runway): Mirelo SFX 2.0 is available in Runway MCP on paid plans. It takes a clip and a description of the desired sound, then generates a separate audio track.
• Qwen3.8-Flash-Next-DFlash2 (Inco AI): Qwen3.8-Flash-Next-DFlash2 is a block-diffusion draft model for speculative decoding with Qwen3.8-Flash-Next. It runs inside a speculative decoding server and is not a standalone language model.
• solidcv (solidSF): solidcv is a pure-Rust computer vision and spatial-computing library that also provides tools for VLM token pruning and mechanistic interpretability. Its repository describes a cloud binary WebSocket API for selected engines, metered in solidSF Credits.
• OpenGCUA Cross-Device SDK (Tongyi-MAI): OpenGCUA Cross-Device SDK provides a programming interface for computers, browsers, and phones. It has adapters for macOS, Android, browsers, and Linux, with Python bindings over a Rust core.
• LiteRT Tensor API (Google): LiteRT Tensor API is a C++ authoring layer for building and optimizing on-device model computation graphs. It supports running graphs through LiteRT on CPU, GPU, or NPU, or directly on CPU with XNNPACK.
• OpenGCUA for macOS (Tongyi-MAI): OpenGCUA for macOS is a menu-bar assistant with a floating task panel, voice, and a bundled harness, SDK, and driver. The repository describes it as a local development build and says public-release validation remains to be completed.
• CloudWatch Omni (AWS): CloudWatch Omni is an AI-powered observability experience that centralizes telemetry, dependency mapping, and troubleshooting across AWS accounts, Regions, and other clouds. It provides a standalone web console with SSO, IDE extensions for local agent instrumentation, and toolkits to query and interact with telemetry via natural language or point-and-click workflows.
• Copilot code review (GitHub): Copilot code review adds a dedicated per-user settings page and more granular automatic-review controls, plus an enterprise-level default review effort that can be inherited by organizations. These changes are generally available across Copilot plans including Business and Enterprise.
• GitHub Copilot (GitHub): GitHub deprecated four models across its Copilot experiences on October 2, 2026. It suggests replacement models and says administrators may need to enable access to alternatives through model policies.
• Copilot CLI (GitHub): Copilot CLI adds whole codebase indexing (WCI) for C++ that builds a persistent symbol index across a project so the Microsoft C++ Language Server can resolve symbols and relationships without repeatedly rediscovering them. The feature is enabled by default, can increase initial build time and memory usage, and the index is reused and dynamically updated thereafter.
• Copilot for JetBrains 1.18.0 (GitHub): Copilot for JetBrains 1.18.0 is a JetBrains IDE plugin update that adds new agent features, organization-shared skills and instructions, and workflow improvements. The release includes public-preview assisted tool approvals, a re-edit message capability, Codex agent plan mode, and controls for MCP tools and the built-in MCP server.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• Whistle adds local fine-tuning, so developers can adapt its on-device speech transcription to a speaker, vocabulary, or new language.
• SupraTTS-1.0-Preview generates English speech locally from text on a CPU, without needing a network connection at runtime.

Whistle can now be tuned for a speaker or new language

Cactus Compute’s Whistle speech recognition model now supports local fine-tuning: developers can adapt it to particular speakers or vocabulary, or teach it a new language. The tuned model can be exported as a .cact file. Whistle runs on-device on a CPU and can transcribe speech in seven languages, add word timestamps, and generate speech embeddings. These features may be useful for apps and devices that need speech processing locally, including phones, wearables, and smart-home devices.

SupraTTS can create English speech without a network connection

SupraTTS-1.0-Preview is an open-weight text-to-speech model that turns English text into audio and is designed to run locally with a Python runtime. It can synthesize speech on a CPU without requiring a network connection at runtime. That may suit developers building tools that need offline English speech generation; the record describes it as a preview release.

The records

• Whistle (Cactus Compute): Whistle is a 16.9 MB speech recognition model that runs on-device on a CPU. It transcribes speech in seven languages and can also return word timestamps and speech embeddings.
• SupraTTS-1.0-Preview (SupraLabs): SupraTTS-1.0-Preview is an open-weight English text-to-speech model with 1,031,754 parameters. It generates 24 kHz mono audio and is designed to run locally with a Python runtime.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• DeepSeek opened a search API that returns real-time, multilingual web results, and released an upgraded model it says improves benchmark scores while reducing runtime and output length.
• Graphicle gives coding agents searchable, cited context from a codebase and can preserve reviewed project memories and past sessions.
• New open-weight models target on-device text classification, financial research, and healthcare tasks; their capabilities and input types differ.

DeepSeek adds web search for apps and updates its Flash model

DeepSeek Search API gives developers access to real-time results from DeepSeek’s search engine, which supports multiple languages and indexes hundreds of billions of web pages. It could let an app or AI assistant retrieve current web information; the record does not specify how results are ranked or what access costs. DeepSeek-V4.1-Flash-1017 is an upgraded version of DeepSeek V4.1 Flash, available through an API, website, and app. DeepSeek reports improved benchmark results alongside shorter task runtimes and outputs, but those are the company’s claims rather than independently described test results.

A separate set of Rust libraries, deepseek-recipe, helps developers format prompts for DeepSeek V4 and V4.1 and parse responses, including streamed replies. It does not run the model or execute tool calls.

Give coding agents cited context from a project

Graphicle indexes code and project materials so coding agents can search them and receive cited excerpts. It can also capture agent sessions and maintain project memories that people have reviewed, giving teams a way to carry decisions and relevant past work into later coding tasks. It is available as a local-first application with a command-line interface, browser app, and MCP server. The record describes its features, not evidence that it improves coding results.

New models focus on classification, finance, and healthcare

GLiClass v4 Edge is a small model designed to run on CPUs and edge devices. Rather than writing text, it scores how well supplied labels or instructions fit a piece of text, which can be used for tasks such as moderation or ranking. Fastino Nemotron 3.5 Super VL Finance is a text-only model tuned for financial reasoning, extracting information, and research grounded in sources. Its published checkpoint does not include the vision encoder, despite “VL” in its name. Fastino Nemotron 3.5 Super VL Healthcare is also a text-only checkpoint, tuned for biomedical questions, clinical calculations, extracting information, and summarizing clinical notes. These are specialized tools for developers, not evidence of clinical performance.

Also on the radar

• Comet 60B is an open-weight model that accepts text, images, and video, and supports configurable reasoning and tool use for locally served applications.
• MOSS-image_experimental offers image-generation model variants for experimentation, but its page warns that the SFT model can produce clear vertical banding artifacts.

The records

• GLiClass v4 Edge (Knowledgator): GLiClass v4 Edge is a 0.4B-parameter model for instruction-following zero-shot classification, built for CPUs and edge devices. It scores supplied labels in one forward pass and generates no text.
• MOSS-image_experimental (Kijai): MOSS-image_experimental is an experimental collection of MOSS-image-related models, including SFT and LoRA variants. The page notes that the SFT model can produce clear vertical banding artifacts and that applying later blocks with negative strength may reduce them.
• Comet 60B (Multiverse Computing): Comet 60B is a compressed multimodal model based on NVIDIA Nemotron 3.5 Super, with 61.1B total parameters. It accepts text, image, and video input, generates text, and supports configurable reasoning and tool calling.
• Fastino Nemotron 3.5 Super VL Finance (Fastino Labs): Fastino Nemotron 3.5 Super VL Finance is a 120B-parameter, 12B-active text model fine-tuned for financial reasoning, extraction, and research. Its published BF16 checkpoint is a merged model that combines the Nemotron 3.5 Super VL language backbone with Fastino’s finance adapter; the vision encoder is not included.
• Fastino Nemotron 3.5 Super VL Healthcare (Fastino Labs): Fastino Nemotron 3.5 Super VL Healthcare is a 120B-parameter, 12B-active model fine-tuned for healthcare and biomedical applications. The published BF16 checkpoint is text-only and combines the Nemotron 3.5 Super VL language backbone with Fastino's healthcare adapter.
• DeepSeek-V4.1-Flash-1017 (DeepSeek): DeepSeek-V4.1-Flash-1017 is an upgraded version of DeepSeek V4.1 Flash, available through the API, official website, and app. DeepSeek reports benchmark improvements alongside reduced task runtime and output length.
• deepseek-recipe (DeepSeek): deepseek-recipe is a set of Rust libraries with Python bindings for encoding and parsing DeepSeek V4 and V4.1 prompts and responses. It supports thinking, tool calls, images, and generation settings, while leaving model inference, tool execution, and HTTP transport to the caller.
• DeepSeek Search API (DeepSeek): DeepSeek opened access to its in-house search API, powered by the DeepSeek search engine. The engine indexes hundreds of billions of web pages, supports multilingual search, and returns real-time results.
• llama.cpp v0.5.0 (ggml-org): llama.cpp v0.5.0 is a new release of the open-source C/C++ inference library and server for running GGML-format models locally and on servers. The release adds performance optimizations, expanded model/format conversion support, server/router robustness improvements, and UI/server feature updates.
• llama.cpp v0.6.0 (ggml-org): llama.cpp v0.6.0 is a release of the open-source C/C++ inference framework. It adds APIs and model support, server and embedding updates, and changes to its Web UI and hardware backends.
• llama.cpp b11149 (ggml-org): llama.cpp b11149 is a pre-release of the open-source C++ inference library that provides local runtimes and build artifacts for running LLMs across many platforms and backends. This release adds a -b/--backend option to test-llama-archs and ships assets and CI attestations for numerous OS/architecture/backend combinations.
• Graphicle (Graphicle AI): Graphicle is a local-first code graph and memory application for coding agents, available as a Rust program with a command-line interface, browser app, and MCP server. It indexes code and project materials, returns cited excerpts, and can capture agent sessions and maintain reviewed project memories.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• IBM’s Granite Switch 4.1 30B combines chat, speech transcription, decision-making, and safety checks behind one endpoint, with adapters that can be switched on when needed.
• Tencent Hunyuan introduced a memory model, a developer toolkit, and a local app for storing and retrieving information across conversations and coding work.
• New tools target practical workflows: coordinating coding agents’ changes with gitai, and letting agents inspect or create audio with earshot.

One model endpoint for chat, speech, decisions, and safety checks

Granite Switch 4.1 30B from IBM brings a base language model together with optional decision and safety components and a built-in speech-to-text encoder. Its adapters can be activated on demand, so an application can use the same endpoint for natural-language responses, option-based decisions, safety-risk detection, and English speech transcription.

Tencent Hunyuan offers three ways to manage AI memory

Tencent Hunyuan has introduced Hy-Memory-30B-A3B, a model designed to decide what from a conversation to retain and how to update persistent memory. Developers can use it as one component in an external memory system; it is not described as a complete memory product on its own. For developers who want a ready-made framework, Hy Memory 2.0 is a Python toolkit with separate memory pipelines for chat, working sessions, and coding. MemHub offers a local-first desktop app and command-line tool for managing conversation histories and coding traces from supported AI applications, with data kept in a local directory.

Tools help coding agents coordinate changes and work with sound

gitai is a version-control system built for coding agents working on the same repository. It supports concurrent changes, change logs, and configurable ways to resolve conflicts, giving developers tools to review and land agent-made updates. earshot lets coding agents inspect audio and video soundtracks through a command-line tool, Python interface, or MCP server. It can also create sounds from code and refine them using an audio critic, which may be useful for developers building or testing audio features.

An image-retrieval model can rank text and visual matches

Nemotron-3.5-Rerank-8B-BF16 from NVIDIA scores how relevant a document is to a text query. Documents can contain text, images, or both, so developers can use it to reorder candidate results in text or vision search systems, including retrieval-augmented generation applications.

Also on the radar

• Krea2 Realism Checkpoint V1 is a downloadable image-generation model for ComfyUI, described as supporting natural-looking skin tones and textures as well as NSFW generation.
• MOSS-Image-SFT-8-steps-dmd is a work-in-progress image model release with weights and instructions for generating images using an eight-step sampling process.
• The Agentic Scientist is an installable research tool that coordinates coding agents through experiment planning, evaluation, and paper writing. Its multi-stage process runs from a task description, but the record does not establish how well it performs research.
• HAST-SEC v0.1 is a locally runnable Chinese-language demo for chatting with a virtual character and issuing supported motion commands. It includes source code and runtime resources, but not a complete training workflow.
• NVIDIA’s NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-GGUF is a draft checkpoint that must be paired with a compatible target model; it proposes text tokens for that model to verify and is not a standalone language model.

The records

• Granite Switch 4.1 30B (IBM): Granite Switch 4.1 30B is a modular language model combining a base model, decision and safety adapters, and an embedded speech-to-text encoder. Its adapters can be activated on demand, and chat, decisions, and transcription can run through one endpoint.
• Nemotron-3.5-Rerank-8B-BF16 (NVIDIA): Nemotron-3.5-Rerank-8B-BF16 is an 8.38B-parameter reranker that scores the relevance of a document to a text query. It accepts text, images, or combined text and images as documents, and is described as supporting multilingual and multimodal retrieval.
• Krea2 Realism Checkpoint V1 (RudySen): Krea2 Realism Checkpoint V1 is a downloadable image-generation checkpoint made by merging a custom Realism Checkpoint with V3 LoRA. It supports NSFW generation and is described as producing natural skin tones and textures.
• MOSS-Image-SFT-8-steps-dmd (lodestones): MOSS-Image-SFT-8-steps-dmd is a work-in-progress distillation of OpenMOSS-Team’s MOSS-Image-SFT image model. The release includes model weights and instructions for loading and sampling with the checkpoint.
• Hy-Memory-30B-A3B (Tencent Hunyuan): Hy-Memory-30B-A3B is a compact long-term-memory writer model with approximately 3B active parameters, designed to decide what information to retain and how to update persistent memory. Its weights are available on Hugging Face, and the model can be deployed with vLLM and called through an OpenAI-compatible API.
• NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr-GGUF (NVIDIA): This is a LiLiCorr speculative-decoding draft checkpoint for the compatible NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF target model. It proposes token candidates that the target verifies and is not a standalone language model.
• Hy Memory 2.0 (Tencent Hunyuan): Hy Memory 2.0 is a Python SDK and cognitive memory engine for LLM agents, released as `hy-memory==2.0.2`. It provides separate memory pipelines for chat, working, and coding tasks.
• gitai (feder-cr): gitai is an open-source, server-centric version control system designed for coding agents, with clients and server-side history operations. It supports concurrent change landing, changeset logs, configurable conflict resolution, and storage using SQLite or PostgreSQL with object storage.
• MemHub (Tencent Hunyuan): MemHub is a local-first desktop and CLI product built on Hy Memory. It brings conversation history and coding traces from user-authorized AI applications into a managed runtime and local data directory.
• earshot (Leonxlnx): earshot is an open-source tool that lets coding agents analyze audio and video soundtracks through a CLI, Python API, or MCP server. It can also synthesize sounds from code and iteratively refine them using its audio critic.
• HAST-SEC v0.1 (武汉光影流明科技有限公司): HAST-SEC v0.1 is a locally runnable Chinese-language interaction demo for controlling continuous character motion. It uses an approximately 0.8B text model and a small motion model, and includes source code and runtime resources but no complete training workflow.
• The Agentic Scientist (SakanaAI): The Agentic Scientist is a public, installable tool that orchestrates coding agents through scientific idea generation, experiments, evaluation, and paper writing. It runs a multi-phase pipeline from a task directory and supports several agent CLIs and models.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• Bespoke-Nimble-9B-v2 from Bespoke Labs: Bespoke-Nimble-9B-v2 is a PEFT LoRA adapter for text-based, schema-conditioned decisions, requiring the separately downloaded Qwen/Qwen3.5-9B base model.
• open_scion from feder-cr: open_scion is an open-source source-control system designed for coding-agent workflows, implementing a server-centric design with normal Git repositories underneath.
• Ternary Bonsai 2 9B from PrismML: Ternary Bonsai 2 9B is a 9-billion-parameter model based on Qwen3.5, using PrismML’s ternary weight representation.

Also on the radar

• Ternary Bonsai 2 4B (PrismML): Ternary Bonsai 2 4B is a 4-billion-parameter model based on Qwen3.5, using PrismML’s ternary weight representation.
• Ternary Bonsai 2 Ternary 9B (PrismML): Ternary Bonsai 2 Ternary 9B is a 9B-class vision-language model with ternary transformer weights, distributed in GGUF formats for llama.cpp.
• Ternary Bonsai 2 2B (PrismML): Ternary Bonsai 2 2B is a 2-billion-parameter model based on Qwen3.5, using PrismML’s ternary weight representation.
• d1-ultra (Liquid AI): d1-ultra is a multimodal decision model that accepts text, images, or both and returns probabilities for yes/no, choice, and score questions in one forward pass.
• Claude Opus 5.5 (Anthropic): Claude Opus 5.5 is a hosted Anthropic model for long-running agentic coding and knowledge work, released with a 1,000,000-token default context window and 128k max output tokens.
• Copilot code review (GitHub): Copilot code review adds a dedicated per-user settings page and more granular automatic-review controls, plus an enterprise-level default review effort that can be inherited by organizations.
• GitHub Copilot (GitHub): GitHub deprecated four models across its Copilot experiences on October 2, 2026.
• Copilot CLI (GitHub): Copilot CLI adds whole codebase indexing (WCI) for C++ that builds a persistent symbol index across a project so the Microsoft C++ Language Server can resolve symbols and relationships without repeatedly rediscovering them.
• Copilot for JetBrains 1.18.0 (GitHub): Copilot for JetBrains 1.18.0 is a JetBrains IDE plugin update that adds new agent features, organization-shared skills and instructions, and workflow improvements.

The records

• Bespoke-Nimble-9B-v2 (Bespoke Labs): Bespoke-Nimble-9B-v2 is a PEFT LoRA adapter for text-based, schema-conditioned decisions, requiring the separately downloaded Qwen/Qwen3.5-9B base model. It scores allowed answer-token logits to provide boolean judgments, categorical choices, ordinal scores, and probability distributions; it does not generate reasoning or free-form answers.
• Ternary Bonsai 2 9B (PrismML): Ternary Bonsai 2 9B is a 9-billion-parameter model based on Qwen3.5, using PrismML’s ternary weight representation. It supports reasoning, visual understanding, and tool-driven workflows in a compact model footprint.
• Ternary Bonsai 2 4B (PrismML): Ternary Bonsai 2 4B is a 4-billion-parameter model based on Qwen3.5, using PrismML’s ternary weight representation. It supports reasoning, visual understanding, and tool-driven workflows in a compact model footprint.
• Ternary Bonsai 2 Ternary 9B (PrismML): Ternary Bonsai 2 Ternary 9B is a 9B-class vision-language model with ternary transformer weights, distributed in GGUF formats for llama.cpp. It supports text and image input and has a 262,144-token context length inherited from its Qwen3.5-9B backbone.
• Ternary Bonsai 2 2B (PrismML): Ternary Bonsai 2 2B is a 2-billion-parameter model based on Qwen3.5, using PrismML’s ternary weight representation. It supports reasoning, visual understanding, and tool-driven workflows in a compact model footprint.
• d1-ultra (Liquid AI): d1-ultra is a multimodal decision model that accepts text, images, or both and returns probabilities for yes/no, choice, and score questions in one forward pass. Liquid AI says it is based on a larger open-weight model with Liquid-specific customizations and is trained to produce more reliable probabilities.
• Claude Opus 5.5 (Anthropic): Claude Opus 5.5 is a hosted Anthropic model for long-running agentic coding and knowledge work, released with a 1,000,000-token default context window and 128k max output tokens. It enforces always-on adaptive thinking and is deployed across multiple cloud platforms and the Claude API with migration guidance for tool and computer toolset changes.
• open_scion (feder-cr): open_scion is an open-source source-control system designed for coding-agent workflows, implementing a server-centric design with normal Git repositories underneath. It provides server-side history and code search, compare-and-swap landing, agent-oriented review and security features, and optional conflict resolution through an LLM provider, deterministic tool, or no resolver.
• Copilot code review (GitHub): Copilot code review adds a dedicated per-user settings page and more granular automatic-review controls, plus an enterprise-level default review effort that can be inherited by organizations. These changes are generally available across Copilot plans including Business and Enterprise.
• GitHub Copilot (GitHub): GitHub deprecated four models across its Copilot experiences on October 2, 2026. It suggests replacement models and says administrators may need to enable access to alternatives through model policies.
• Copilot CLI (GitHub): Copilot CLI adds whole codebase indexing (WCI) for C++ that builds a persistent symbol index across a project so the Microsoft C++ Language Server can resolve symbols and relationships without repeatedly rediscovering them. The feature is enabled by default, can increase initial build time and memory usage, and the index is reused and dynamically updated thereafter.
• Copilot for JetBrains 1.18.0 (GitHub): Copilot for JetBrains 1.18.0 is a JetBrains IDE plugin update that adds new agent features, organization-shared skills and instructions, and workflow improvements. The release includes public-preview assisted tool approvals, a re-edit message capability, Codex agent plan mode, and controls for MCP tools and the built-in MCP server.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• Indra OCR can turn PDFs and images into structured text locally, including on a CPU.
• Dwight 4B 1.0 is built to classify sales-call outcomes or produce timestamped insights with supporting quotes.
• Grok Voice 3.0 is now available through an API for speech-to-speech applications.
• Claude Platform added a beta option for tracing requests and sessions, alongside changes for evaluating and monitoring agent tool calls.

Extract structured text from PDFs and images on a CPU

Indra OCR, from Datalab, converts PDFs and images into Markdown, plain text, or JSON with bounding boxes. It runs locally on a CPU, so it could be useful for extracting and organizing documents without relying on a hosted service. It also identifies lines and layout blocks and provides their reading order, which can help preserve a document’s structure.

A model for classifying sales calls with evidence

Dwight 4B 1.0, from Zime AI, analyzes sales-call transcripts and can either make a one-token decision with a confidence score or return structured insights. Those insights can include timestamps and verbatim quotes, making it easier to connect a classification to the relevant part of a call.

Speech-to-speech voice experiences through an API

Grok Voice 3.0 is available through an API with Speech to Speech. Developers can use it to build voice experiences that take in speech and respond with speech; the API is unchanged from Grok Voice 2.0, and pricing is the same.

New ways to evaluate and monitor Claude agent sessions

The Claude Platform update adds an “auto” permission mode that lets the server evaluate Managed Agents tool calls, with evaluation details reported in agent events. The ant CLI also gained a beta command for connecting a terminal to a session, plus an option to open the session viewer locally. OpenTelemetry export for tracing Messages API requests and Managed Agents sessions is available in beta. This release is dated September 10, 2026, so it is an earlier update rather than a release from October 14.

Also on the radar

• Fish Audio Drama 3 supports text-to-speech for multi-character scenes and long-form content, with voice cloning and up to 16 speaker tags.
• QUAN-Transcribe-1.0 can transcribe Vietnamese audio offline, either from a complete recording or as a stream.

The records

• Indra OCR (Datalab): Indra OCR is a 1M-parameter OCR model that converts PDFs and images into Markdown, plain text, or JSON with bounding boxes. It runs on CPU using ONNX Runtime and provides line detection, layout blocks, and reading order.
• HydraFusion Max (GitHub): HydraFusion Max is a separately selectable option built on HydraFusion's architecture and orchestration techniques. It is tuned for frontier capability on more demanding tasks, applying higher reasoning effort where depth and thoroughness matter most.
• Grok Voice 3.0 (xAI): Grok Voice 3.0 is available through the API with Speech to Speech. It is priced the same as Grok Voice 2.0, and the API is unchanged.
• Supra2.5-Ultra-Preview-001 (SupraLabs): Supra2.5-Ultra-Preview-001 is a 230-million-parameter base model with a Llama architecture, trained on about 40.4 billion tokens. It is available to download and use with Transformers, and the publisher cautions that it may generate incorrect or biased text and has not been aligned or evaluated for safety.
• Dwight 4B 1.0 (Zime AI): Dwight 4B 1.0 is a 4-billion-parameter model for analyzing sales-call transcripts, based on Qwen3-4B-Instruct-2507. It produces one-token decisions with calibrated confidence or structured JSON insights with timestamps and verbatim quotes.
• Ling-3.1-flash (inclusionAI): Ling-3.1-flash is a hybrid reasoning model with 560B total parameters, 25B active parameters per token, and a 1M-token context window. Its model weights and instructions for running inference with SGLang and vLLM are publicly provided.
• Amos 1 large (Mukaddam Zaid): Amos 1 large is a 27B decision model that answers typed questions with probabilities for each allowed answer, without generating free-form text. It accepts text, JSON, or photos and is available through a hosted API or for self-hosting with its published weights and serving package.
• UHR-BAT2 (RL-MIND): UHR-BAT2 is a full-weight fine-tune of Qwen3-VL-8B-Instruct for remote-sensing visual question answering. Its released BF16 safetensors checkpoint can be loaded with Transformers without custom remote code.
• Fish Audio Drama 3 (Fish Audio): Fish Audio Drama 3 is a text-to-speech model designed for multi-character scenes and long-form content. It supports voice cloning, natural-language control, and generation with up to 16 speaker tags.
• Lumma-Fev-27B (FrontiersMind): Lumma-Fev-27B is a full-weight fine-tune of Qwen3.8-27B that returns probabilities for typed decision questions instead of generating text. It supports English, Hindi, Tamil and Kannada, and can also accept images as input.
• QUAN-Transcribe-1.0 (qnguyen3): QUAN-Transcribe-1.0 is a Vietnamese automatic speech recognition model built on EmbeddingGemma 2. It supports offline transcription and streaming inference.
• Claude Platform (Anthropic): Anthropic updated the Claude Platform to let the server automatically evaluate Managed Agents tool calls via a new "auto" permission mode and to report evaluation metadata in agent events. The ant CLI (v1.32.0) gained a beta sessions connect command to attach a terminal to Managed Agents sessions and a --web option to open the session viewer locally.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• LFM2.5-24B-A2B is an open-weight model designed for reasoning and tool use; its base checkpoint is intended for people who want to fine-tune a model for a specific task.
• Mizan v3 3B can classify records, score them, or estimate probabilities for allowed answers, including choices and yes-or-no questions.
• A Korean speech-recognition model fine-tuned from Qwen3-ASR-1.7B is publicly available for transcribing audio.

An open model for tool-using agents, plus a checkpoint to fine-tune

LFM2.5-24B-A2B is a text model designed for reasoning and tool use. Developers can run it with Transformers, vLLM, or SGLang, or fine-tune it for a particular use. Liquid AI also released LFM2.5-24B-A2B-Base, a pretrained starting point intended for substantial fine-tuning, such as adapting an assistant to a specific field or proprietary data. Both records describe a 131,072-token context length; that specification alone does not show how well either model handles long inputs.

Mizan can sort records or score them against rules you provide

Mizan v3 3B, from Nimruz, takes text or structured JSON records and a typed question, then returns probabilities for the allowed answers. It supports choices, yes-or-no questions, and scoring, and can apply policies or rubrics included in the record. That could help developers route incoming records, rate severity, or classify intent. It is a Persian-first model, and the supplied record describes it as available with open weights, a CLI, and an API.

Also on the radar

• A publicly available fine-tune of Qwen3-ASR-1.7B is intended to transcribe Korean speech from audio files.
• VegaML is a Python package for making decisions from text or images, returning choices, scores, or probabilities alongside information that can indicate uncertainty or allow it to abstain.

The records

• Mizan v3 3B (Nimruz): Mizan v3 3B is a Persian-first decision model that takes text or JSON records and typed questions, then returns probabilities for each allowed answer in one forward pass. It supports choice, yes/no, and scoring questions, and can apply policies and rubrics included in the record.
• LFM2.5-24B-A2B (Liquid AI): LFM2.5-24B-A2B is a 24B-parameter text-only reasoning model with 2.45B active parameters, post-trained for agentic workloads. It supports tool calling, has a 131,072-token context length, and is released as a model checkpoint for use with Transformers, vLLM, and SGLang.
• qwen3_asr_1.7b_ko (TeamUNIVA): This is a Korean speech recognition fine-tune of Qwen3-ASR-1.7B, trained with two rounds of supervised fine-tuning and reinforcement learning. The model checkpoint and usage instructions are publicly available on Hugging Face.
• LFM2.5-24B-A2B-Base (Liquid AI): LFM2.5-24B-A2B-Base is a 24B-parameter, text-only pretrained checkpoint with 2.45B active parameters and a 131,072-token context length. Liquid AI describes it as part of a family designed for on-device deployment and recommends it for tasks requiring substantial fine-tuning.
• VegaML (NandhaKishorM): VegaML is a Python package for making typed decisions from text or images using a frozen language model and a physics engine. It returns choices, scores, or boolean probabilities with calibrated probabilities, conformal answer sets, and an abstain flag.
• Copilot code review (GitHub): Copilot code review adds a dedicated per-user settings page and more granular automatic-review controls, plus an enterprise-level default review effort that can be inherited by organizations. These changes are generally available across Copilot plans including Business and Enterprise.
• GitHub Copilot (GitHub): GitHub deprecated four models across its Copilot experiences on October 2, 2026. It suggests replacement models and says administrators may need to enable access to alternatives through model policies.
• Copilot CLI (GitHub): Copilot CLI adds whole codebase indexing (WCI) for C++ that builds a persistent symbol index across a project so the Microsoft C++ Language Server can resolve symbols and relationships without repeatedly rediscovering them. The feature is enabled by default, can increase initial build time and memory usage, and the index is reused and dynamically updated thereafter.
• Copilot for JetBrains 1.18.0 (GitHub): Copilot for JetBrains 1.18.0 is a JetBrains IDE plugin update that adds new agent features, organization-shared skills and instructions, and workflow improvements. The release includes public-preview assisted tool approvals, a re-edit message capability, Codex agent plan mode, and controls for MCP tools and the built-in MCP server.
• Amazon Connect Customer (AWS): Amazon Connect Customer now supports agent-to-agent collaboration, allowing specialized AI agents to be brought into a live customer interaction. The launch expands the open A2A protocol to support voice streaming and interaction-level session continuity and provides unified observability, guardrails, and escalation controls across collaborating agents.
• Copilot app (Local sandboxing) (GitHub): Local sandboxing for the Copilot app provides per-project policies that limit filesystem, network, and credential access for local repository and working-tree sessions. The feature is available in public preview and is configurable in the app per project or for individual sessions with an in-session command.
• Facebook Plus (Meta): Facebook Plus is a paid subscription that adds personalization and other features to Facebook. The announcement says it includes custom app icons, super reactions and additional Reels insights.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• Meta introduced Muse Wonder, a model it says can handle personal and professional tasks such as interpreting documents and finding flaws in code.
• Two models turn supplied questions into structured decisions: LemonDecider can choose, score, or answer yes/no with an unknown option, while EmbeddingGemma 2 · typed decisions returns typed answers with probabilities.
• OppaiOracle experimental-X1.1 can suggest tags for anime and illustration images, with scores that people can review before using them.

Muse Wonder is aimed at document work, workflows, and coding

Muse Wonder, from Meta, is a new model for tasks that can involve several steps, including interpreting documents, handling workflows, and finding flaws in code. Meta says it performs at the leading edge on personal and professional tasks, including expert fields and coding; the announcement does not provide details here to independently compare that claim.

Two models turn questions into structured decisions

LemonDecider and EmbeddingGemma 2 · typed decisions are designed to answer questions in a fixed format rather than generate an open-ended response. LemonDecider can choose from supplied options, make a yes-or-no judgment with an explicit unknown option, or score ordered descriptions; it can also accept images. EmbeddingGemma 2 · typed decisions takes a workflow state and a changeable question schema, then returns answers with probabilities. These may be useful when an application needs consistent labels or scores—for example, deciding whether a record should be reviewed by a person.

A new model suggests tags for anime and illustration images

OppaiOracle experimental-X1.1 identifies possible tags for anime and illustration images, returning tag names with scores. Its release includes scripts and a vocabulary of nearly 15,000 tags. The scores let someone review suggestions before adding tags to a dataset; the model is intended for image tagging, not for deciding automatically which labels are correct.

Also on the radar

• MixLID identifies language spans in mixed-language text and predicts the main language, supporting 111 language labels.
• AuroraIMG-6M generates small images from short English captions and can run on a CPU; MiniCPM-V 4.7 1B DSpark is a draft model for speculative decoding with MiniCPM-V 4.7 1B.

The records

• MiniCPM-V 4.7 1B DSpark (OpenBMB): MiniCPM-V 4.7 1B DSpark is a draft model trained for MiniCPM-V 4.7 1B. The source links to its separately released model artifact.
• OppaiOracle experimental-X1.1 (Grio43): OppaiOracle experimental-X1.1 is an experimental multi-label tagger for anime and illustration images, released as FP32 ONNX and safetensors weights. It uses 448×448 inputs and returns tag names with sigmoid scores; the release includes inference scripts and a 14,994-tag vocabulary.
• Muse Wonder (Meta): Muse Wonder is a model in Meta’s Muse Wonder series, built on the groundwork laid with Muse Spark. Meta says it delivers state-of-the-art performance on personal and professional agentic tasks, including expert domains and coding.
• LemonDecider (inclusionAI): LemonDecider is a typed decision model derived from Qwen3.5-4B, designed to answer choice, noul, and score questions in a single forward pass with an explicit unknown candidate. It uses a 255-way decision readout rather than an autoregressive generation head and optionally accepts images.
• MixLID (nlpai-lab): MixLID is a token-classification model for identifying languages in monolingual and mixed-language text. It supports 111 language labels and provides token-level spans as well as a predominant-language prediction for an input.
• EmbeddingGemma 2 · typed decisions (mrm8488): EmbeddingGemma 2 fine-tuned for typed decisions takes a workflow state and a schema of questions and returns typed answers with probabilities. It supports choice, score, and probabilistic yes/no question types, and the schema can be changed without retraining.
• Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated (huihui-ai): This is an uncensored version of nerkyor’s Qwen3.8-27B-Coder390-EfficientThink model, created with abliteration. The listing describes it as a crude proof of concept and says only layers 2 to 52 have been ablated.
• AuroraIMG-6M (AuroraAI-Research): AuroraIMG-6M is a 6.82-million-parameter text-to-image model that generates 64×64 RGB images from short English captions. Its weights are available under the Apache 2.0 license, and the model can run on a CPU.
• Mistral Small 4 (Mistral): Mistral says it is releasing Mistral Small 4 for developers, researchers, and organizations. The announcement describes it as part of its commitment to open models but gives no model specifications or access details here.
• Voxtral TTS (Mistral): Voxtral TTS is a 4B-parameter text-to-speech model that generates expressive speech in nine languages. It supports low-latency generation and voice adaptation from a short reference, and is available through an API, Mistral Studio, Le Chat, and as open weights.
• Workflows (Mistral): Workflows is an orchestration layer for enterprise AI processes, available in public preview as part of Studio. Developers write workflows in Python, and teams can trigger them through Le Chat, with execution history and auditing in Studio.
• RSI-Anything (Gen-Verse): RSI-Anything is a domain-general framework that combines world-model predictions, interaction feedback, external evaluation, and learning in a recursive improvement cycle. Its public repository provides a Python command-line workflow and task implementations for code repair, GUI interaction, laboratory planning, reaction optimization, and immune science.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• AesCode-8B and AesCode-32B generate editable slides, posters, dashboards, and other visual documents from prompts, with an optional reference image for guidance.
• JEV-35B and d3 answer decision questions with choices, yes-or-no answers, or scores and probabilities, rather than generating chat responses.
• Several llama.cpp releases bring updates to local model running, but the records repeat the same pre-release note across different versions and need editorial review.

Generate editable slides and dashboards from a prompt

AesCode-8B generates visual documents such as slides, posters, dashboards, and reports as structured HTML and CSS. You can provide a reference image to guide the layout and style; the output includes inspectable tables and chart specifications. AesCode-32B is a companion model evaluated for generating information-rich visual artifacts using an image-generated reference. Both models are listed as available on the web with open weights.

Models that choose, classify, or score instead of chatting

JEV-35B can classify requests, score items, or choose among options, including when given images. It returns probabilities and is not designed to chat or generate text. d3 handles choice, yes-or-no, and scoring questions about text, images, or video, returning probabilities without generating text. JEV-35B-NVFP4 is a version of JEV-35B described as requiring less GPU memory than its bf16 counterpart.

Three llama.cpp entries describe updates to local model running

llama.cpp v0.5.0 describes updates to a library and server for running models locally or on servers, including performance and format-conversion changes. llama.cpp v0.6.0 lists API, model-support, server, embedding, interface, and hardware-backend updates. llama.cpp b11149 is marked as a pre-release and adds a backend option for architecture testing, alongside build assets for many operating systems, processor types, and backends. The records may be useful to developers running or serving supported models, but their repeated release note needs checking.

Also on the radar

• PiDurableKit is a Swift package for building AI-agent apps across Apple platforms, with conversations, tasks, tool use, and storage designed to recover after interruptions or crashes.
• Bespoke-Nimble-9B-v3 is an adapter for answering multiple-choice or yes-or-no questions with probabilities based on supplied context.

The records

• AesCode-32B (Microsoft): AesCode-32B is identified as a companion model to AesCode-8B and is evaluated for generating information-rich visual artifacts. The source reports results for it using an image-generated reference and says it was trained with the same recipe as AesCode-8B, with a longer run.
• AesCode-8B (Microsoft): AesCode-8B generates information-rich visual artifacts as structured HTML/CSS, including slides, posters, dashboards, and reports. It can use a reference image for layout and style guidance, and its output includes inspectable HTML tables and ECharts specifications.
• JEV-35B (AutoTrust AI): JEV-35B is an open-weights System 1 decision model built on Qwen3.5-35B-A3B. It selects among options, returns calibrated probabilities, and supports image inputs, but does not chat or generate text.
• d3 (vLLM Semantic Router): d3 is a 27B multimodal decision model that accepts text or JSON and can also process images and videos. It answers choice, yes/no, and scoring questions in one call, returning probabilities without generating text.
• JEV-35B-NVFP4 (AutoTrust AI): JEV-35B-NVFP4 is an NVFP4 version of AutoTrust AI’s System 1 decision model, with its routed experts in NVFP4. It supports the same serve.sh and API as JEV-35B and is described as requiring less GPU memory than the bf16 version.
• Bespoke-Nimble-9B-v3 (Bespoke Labs): Bespoke-Nimble-9B-v3 is a LoRA adapter for Qwen3.5-9B. Given a context and typed choice or yes-or-no questions, it returns a probability for each option.
• PiDurableKit (finnvoor): PiDurableKit is a Swift package for building durable AI agents on iPhone, iPad, Mac, Apple TV, and Vision Pro, powered by pi-durable running in JavaScriptCore. It provides durable conversations, tool execution, tasks, storage, and model-provider integrations through a Swift API.
• llama.cpp v0.5.0 (ggml-org): llama.cpp v0.5.0 is a new release of the open-source C/C++ inference library and server for running GGML-format models locally and on servers. The release adds performance optimizations, expanded model/format conversion support, server/router robustness improvements, and UI/server feature updates.
• llama.cpp b11149 (ggml-org): llama.cpp b11149 is a pre-release of the open-source C++ inference library that provides local runtimes and build artifacts for running LLMs across many platforms and backends. This release adds a -b/--backend option to test-llama-archs and ships assets and CI attestations for numerous OS/architecture/backend combinations.
• llama.cpp v0.6.0 (ggml-org): llama.cpp v0.6.0 is a release of the open-source C/C++ inference framework. It adds APIs and model support, server and embedding updates, and changes to its Web UI and hardware backends.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• Vega’s 0.8B and 4B models turn text or images into choices, yes/no answers, or scores, and can run locally.
• Mistral released an Arabic speech-to-text model that produces transcripts as audio arrives, including for dialects and code-switching.
• GitHub added finer controls for Copilot code reviews and made local sandboxing generally available in several Copilot tools; the Copilot app’s sandboxing remains in public preview.

Two local models can classify, route, or score text and images

Vega 0.8B and Vega 4B are locally runnable models that turn text or images into a choice, a yes-or-no answer, or a score. They’re intended for tasks such as sorting requests, screening messages, or making decisions about long documents and images. Both list a context length of 73,728 tokens. The records describe their capabilities but do not provide comparative results between the two sizes.

Arabic speech can be transcribed as it arrives

Voxtral Mini 4B Realtime Arabic from Mistral converts incoming speech into text in real time. It is designed for Arabic dialects and Modern Standard Arabic, including speech that switches between Arabic and other languages—useful for live captions or voice interfaces.

Copilot adds review controls and local access limits

Copilot code review now has per-user settings and more detailed controls for automatic reviews. Organizations can also inherit an enterprise default for review effort; the changes are generally available across Copilot plans, including Business and Enterprise. GitHub says local sandboxing is generally available in Copilot CLI, the Copilot app, and VS Code sessions using Agent Host. It lets users limit an agent’s access to files, networks, and credentials. The Copilot app’s own per-project sandbox controls are also described as being in public preview, so availability depends on the specific feature and tool.

Also on the radar

• Iris-3B generates images from prompts and has fine-tuned weights for estimating depth from photos, restoring images, and upscaling them.
• AlphaProtein Novo is an open-weights model for designing enzyme structures and sequences around a catalytic motif and ligand context; its repository includes optional downstream design and evaluation steps.

The records

• ForgePlex-M2.1-10M (ForgeWorks): ForgePlex-M2.1-10M is a 9.99-million-parameter decoder-only language model with a 1,024-token context. Its architecture includes GQA, RoPE, RMSNorm, SwiGLU, attention output gates, and refresh gates.
• Iris-3B (speridlabs): Iris-3B is a 3-billion-parameter pixel-space model that generates images from text prompts at 1024 pixels. Fine-tuned weights in the repository also support monocular depth estimation and image restoration and upscaling.
• Vega 0.8B: public intents (nandakishorm): Vega 0.8B is a locally runnable decision model with a frozen Qwen3.5-0.8B backbone and a trained decision engine. It supports text and image inputs, typed choice, boolean, and score answers, and a context length of 73,728 tokens.
• Voxtral Mini 4B Realtime Arabic (Mistral): Voxtral Mini 4B Realtime Arabic is a streaming speech-to-text model fine-tuned from Voxtral-Mini-4B-Realtime-2602. It is designed for Arabic dialects and Modern Standard Arabic, including code-switching, and produces text as audio arrives.
• Bolmo (Ai2): Bolmo is a language-model approach that operates over the bytes representing characters rather than chunks from a fixed subword vocabulary. Ai2 released additional checkpoints so researchers can experiment with the architecture.
• AlphaProtein Novo (Google): AlphaProtein Novo is a generative diffusion model for de novo enzyme design using structural motif scaffolding. Its repository provides a pipeline that can combine the model with optional sequence redesign, structure prediction, and evaluation.
• Vega 4B: public intents (nandakishorm): Vega 4B is a locally runnable decision model with a frozen Qwen3.5-4B backbone and a trained decision engine. It supports text and image inputs, typed choice, boolean, and score answers, and a context length of 73,728 tokens.
• Copilot code review (GitHub): Copilot code review adds a dedicated per-user settings page and more granular automatic-review controls, plus an enterprise-level default review effort that can be inherited by organizations. These changes are generally available across Copilot plans including Business and Enterprise.
• GitHub Copilot (GitHub): GitHub deprecated four models across its Copilot experiences on October 2, 2026. It suggests replacement models and says administrators may need to enable access to alternatives through model policies.
• Copilot CLI (GitHub): Copilot CLI adds whole codebase indexing (WCI) for C++ that builds a persistent symbol index across a project so the Microsoft C++ Language Server can resolve symbols and relationships without repeatedly rediscovering them. The feature is enabled by default, can increase initial build time and memory usage, and the index is reused and dynamically updated thereafter.
• Copilot for JetBrains 1.18.0 (GitHub): Copilot for JetBrains 1.18.0 is a JetBrains IDE plugin update that adds new agent features, organization-shared skills and instructions, and workflow improvements. The release includes public-preview assisted tool approvals, a re-edit message capability, Codex agent plan mode, and controls for MCP tools and the built-in MCP server.
• Copilot app (Local sandboxing) (GitHub): Local sandboxing for the Copilot app provides per-project policies that limit filesystem, network, and credential access for local repository and working-tree sessions. The feature is available in public preview and is configurable in the app per project or for individual sessions with an in-session command.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• Youtu-Parsing-Omni turns documents, images, charts, audio, and video into structured information such as text, tables, transcripts, and reports.
• Antalia-2 Mini offers a local option for turning Turkish text into speech, with model weights and code available under Apache-2.0.
• Project Maya has three compact versions of GLM-5.3-Flash for local use; their memory needs and reported accuracy differ.
• Darwin-27B-ZTC-v2 scores choices and yes-or-no questions without generating text, which could suit decision steps in a workflow.
• Qwen-Image-2.1-Turbo is a checkpoint for generating and editing images with Diffusers.

Turn documents, charts, audio, and video into structured information

Youtu-Parsing-Omni accepts documents, images, charts, geometry figures, audio, and video, then produces structured JSON based on the task prompt. That output can include text, tables, captions, transcripts, or reports, making the model relevant to people building tools that need to extract information from different kinds of media.

Generate Turkish speech locally from text

Antalia-2 Mini is a small text-to-speech model that produces Turkish speech in one synthetic male voice. It includes Turkish text normalization and can be used locally through Python or a command-line interface; its weights and code are released under Apache-2.0.

Three local versions of GLM-5.3-Flash trade memory needs against reported accuracy

Project Maya has released three quantizations of GLM-5.3-Flash for computers with limited combined RAM and GPU memory. Maya-S is reported to retain 97.9% of the full model’s accuracy on the listed zero-shot tasks, while Maya-L is reported to retain 99.2% and pick the same next token 90% of the time. Maya-S24 is reported to use about 1.5 GB less GPU memory than Maya-S and decode 14% faster in a benchmark on one 24 GB Tesla V100.

• Maya-S: 96.5 GB.
• Maya-S24: 94.7 GB.
• Maya-L: 156.3 GB.

A model that scores decisions without generating text

Darwin-27B-ZTC-v2 takes a state and a typed question, then returns probabilities for the available options in one pass without generating tokens. It can score yes-or-no decisions, choose among labels, or rate options against an ordered set of criteria—for example, as one step in a control or workflow decision.

Generate and edit images with Qwen-Image-2.1-Turbo

Qwen-Image-2.1-Turbo is an accelerated checkpoint for generating images from text prompts and editing images. It uses the same 7B visual generation architecture as Qwen-Image-2.1 and loads with the Diffusers QwenImage21Pipeline.

Also on the radar

• Amazon Bedrock Managed Agents adds a setting that can return a human-readable summary of a model’s reasoning alongside its response.
• Copilot code review now has per-user settings and more detailed controls for automatic reviews, with an enterprise default review effort that organizations can inherit.

The records

• Maya-L (mw00): Maya-L is Project Maya's 156.3 GB quantization of GLM-5.3-Flash, designed for computers with a large combined pool of RAM and GPU memory. The source reports that it retains 99.2% of the FP8 model's zero-shot accuracy and picks the same next token 90% of the time.
• Maya-S24 (mw00): Maya-S24 is a 94.7 GB quantization of GLM-5.3-Flash with 2-bit routed experts and 4-bit attention and shared experts. The source reports that it uses about 1.5 GB less GPU memory than Maya-S and decodes 14% faster on one 24 GB Tesla V100 in its benchmark.
• Maya-S (mw00): Maya-S is Project Maya's 96.5 GB compact quantization of GLM-5.3-Flash, made for computers with a smaller pool of RAM and GPU memory. The source reports that it retains 97.9% of the full FP8 model's accuracy on the listed zero-shot tasks.
• Antalia-2 Mini (cloud0day3): Antalia-2 Mini is a small Turkish text-to-speech model with one synthetic male voice and 7.62 million parameters. It produces 48 kHz audio, includes Turkish text normalization, and is released with model weights and code under Apache-2.0.
• Youtu-Parsing-Omni (Tencent Hunyuan): Youtu-Parsing-Omni is a 5B omni-modal parsing model that accepts documents, images, charts, geometry figures, audio, and video. It produces structured JSON covering perception and cognition tasks, with outputs selected by a task prompt.
• Qwen-Image-2.1-Turbo (Alibaba Qwen): Qwen-Image-2.1-Turbo is an accelerated checkpoint for text-to-image generation and image editing. It uses the same 7B visual generation architecture as Qwen-Image-2.1 and loads with the Diffusers QwenImage21Pipeline.
• Darwin-27B-ZTC-v2 (FINAL-Bench): Darwin-27B-ZTC-v2 is a zero-token decision model that takes a state and a typed question and returns probabilities for each option. It uses one forward pass per question and generates no tokens.
• Amazon Bedrock Managed Agents (AWS): Amazon Bedrock Managed Agents is built on a customized version of OpenAI's Agents API and integrated with AWS resources. It manages agent state, tool use, code execution, and multi-step coordination, with durable sessions, reusable skills, and support for MCP servers.
• Amazon Connect Customer (AWS): Amazon Connect Customer now supports agent-to-agent collaboration, allowing specialized AI agents to be brought into a live customer interaction. The launch expands the open A2A protocol to support voice streaming and interaction-level session continuity and provides unified observability, guardrails, and escalation controls across collaborating agents.
• Copilot code review (GitHub): Copilot code review adds a dedicated per-user settings page and more granular automatic-review controls, plus an enterprise-level default review effort that can be inherited by organizations. These changes are generally available across Copilot plans including Business and Enterprise.
• Brewery (Empero): Brewery is a guided console agent for fine-tuning language and image models. It guides users through model and data selection, training, testing, export, and publishing.
• RoboHarness (BinceQu): RoboHarness is an open-source robotic control harness that lets LLM agents operate embodied tasks using visual clicks, tracked keypoints, and geometric constraints. It includes a robot interface, task runner, and evaluation pipeline for the BEHAVIOR simulator.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• LightOnOCR-3-4B and LightOnOCR-3-0.8B can extract document text alongside visual structure such as bounding boxes and chart data.
• AQuA looks for recurring failures in production AI agents, while OSS Scanner offers enrolled open-source projects periodic security scans.
• NeuDecide turns speech directly into tool calls, and Phocinae-Largha-150M-v1 handles structured decisions such as yes/no or pick-one questions.
• Qwen3.8-Max is available in the US (Virginia) region with a native 1M-token context window; ML Drift offers developers GPU tools for on-device AI inference.

Extract document layout and chart data, not just text

LightOnOCR-3-4B and LightOnOCR-3-0.8B can transcribe document text and identify visual structure, including labeled bounding boxes, image descriptions, and chart data. That could help developers build tools for processing forms, receipts, tables, or charts from document images; both releases have open weights.

Find recurring agent failures or scan open-source code for security issues

AQuA (Ambient Quality Agent) samples production agent sessions, checks for failures, groups related findings, and tracks issues across runs. It can investigate a failure against a fixed source snapshot and suggest code edits, but it does not apply them or open pull requests. OSS Scanner is an opt-in service for enrolled open-source projects that provides periodic security scans and reports with an explanation, a proof of concept, and a suggested fix where available. The reports are sent without human review, so maintainers will need to assess the findings themselves.

Turn spoken commands into tool calls, or make simple decisions locally

NeuDecide converts speech directly into structured calls to tools, without first producing a transcript. Its support for tool definitions written as JSON schemas makes it relevant to applications that need spoken commands to trigger actions, including on-device applications such as robot control. Phocinae-Largha-150M-v1 is designed for structured decisions rather than chat: it can answer yes/no or pick-one questions and score options, and is designed to run locally in one pass.

Also on the radar

• Qwen3.8-Max is available in the US (Virginia) region with a native 1M-token context window and support for text, visual input, and function calls.
• ML Drift is an open-source GPU compute engine for on-device AI inference, with desktop previews for Windows and Linux; Kandinsky 6.0 Video Super-Resolution can upscale Kandinsky 6.0 video output to Full HD.

The records

• LightOnOCR-3-4B (LightOn): LightOnOCR-3-4B is the largest model in the LightOnOCR-3 family, designed for OCR with layout and visual understanding. It transcribes document text and can produce labeled bounding boxes, image descriptions, and structured chart data.
• LightOnOCR-3-0.8B (LightOn): LightOnOCR-3-0.8B is the lightest model in LightOn's LightOnOCR-3 family and is available as an open-weight release. It transcribes document text and can also identify visual elements with bounding boxes, describe images, and extract chart data.
• GPT-6.1 Sol (OpenAI): GPT-6.1 Sol is a model for coding, professional work, computer use, and multi-step workflows. OpenAI says it improves on GPT-6 Sol and approaches GPT-6 Astra on several evaluations at substantially lower cost.
• NeuDecide (Neuphonic): NeuDecide is a compact voice-action model that converts speech directly into function calls without producing an intermediate transcript. It accepts tool definitions as JSON schemas and runs using ONNX Runtime.
• Phocinae-Largha-150M-v1 (Phocinae): Phocinae-Largha-150M-v1 is a 144.3M-parameter bilingual model for structured decisions, such as yes/no, pick-one, and score questions about a state. It is not a chat model and is designed to run locally in a single forward pass.
• Qwen3.7-Flash (Alibaba Qwen): Qwen3.7-Flash is available in the US (Virginia) region. It supports text generation, deep thinking, visual understanding, Function Calling, structured output, and context caching.
• Qwen3.8-Max (Alibaba Qwen): Qwen3.8-Max is available in the US (Virginia) region with a native 1M-token context window. It supports text generation, deep thinking, visual understanding, Function Calling, structured output, and context caching.
• Kandinsky 6.0 Video Super-Resolution (KandinskyLab): Kandinsky 6.0 Video Super-Resolution is a plug-in super-resolution model for Kandinsky 6.0 video output. It raises output resolution to Full HD (1920×1080).
• Darwin-27B-ZTC (FINAL-Bench): Darwin-27B-ZTC is a zero-token decision model that returns probability distributions over options for typed questions. The model and inference code are available for download, and the listed weights are BF16.
• ML Drift (Google): ML Drift is an open-source, cross-platform GPU compute engine for on-device AI/ML inference, available as a standalone library and as the core GPU acceleration engine within LiteRT. It supports GPU backends including OpenGL ES, OpenCL, Metal, and WebGPU, with desktop previews for Windows and Linux.
• AQuA (Ambient Quality Agent) (Google): AQuA is an ambient quality agent that samples production agent sessions, reviews them for failures, clusters and verifies findings, and tracks issues across runs. It can analyze failures against an immutable source snapshot and propose code-anchored edits, but it does not apply edits or open pull requests.
• OSS Scanner (Anthropic): OSS Scanner is an opt-in service that gives enrolled open-source projects periodic security scans from Anthropic’s most capable models, free of charge. Reports include a proof of concept, an explanation, and a suggested fix where available, and are sent without human review.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• Liquid AI introduced decision models that can classify or route text and images without generating a full written response; one also accepts speech.
• The Developer Knowledge API lets developers search Google documentation and build answers grounded in it.
• ChatGPT Space gives teams a shared area where ChatGPT can organize work and build on shared knowledge.

Liquid AI’s decision models classify inputs without writing a response

d1-omni-600M and d1-3B are models designed to answer a specific question about an input in one pass, rather than generating a stream of text. That makes them suited to tasks such as sorting support tickets, routing commands, or scoring images. d1-omni-600M accepts text, images, and speech; d1-3B accepts text and images and is available as open weights for local use.

Search Google developer documentation through an API

Google’s Developer Knowledge API gives software access to documentation for Google Cloud, Firebase, and Android. It can search by keywords or meaning, retrieve documents, and answer questions using that documentation as its basis. Developers can use it in tools or applications that need to find current Google documentation and produce answers grounded in those sources.

ChatGPT Space gives teams a shared place to work with AI

ChatGPT Space is a shared team area where teammates and ChatGPT can build on common knowledge. ChatGPT can organize the space according to instructions, and can compose responses with text, visuals, and interactive elements. It is available on desktop and web for eligible plans, with some page functions also available on mobile.

Also on the radar

• Claude Haiku 5.5 is aimed at high-volume, latency-sensitive work and supports long inputs and extended outputs. Claude Sonnet 5.5 has a lower price for reading cached prompts; other prices are unchanged.
• Playground is an experimental browser platform for making and sharing games from text prompts, with multiplayer and leaderboards in select genres.

The records

• d1-omni-600M (Liquid AI): d1-omni-600M is a 587M-parameter decision model built on LFM2.5-Encoder-350M. It answers named questions about text, images, or speech in a single pass, returning typed answers without generating output tokens.
• d1-3B (Liquid AI): d1-3B is an open-weight decision model that produces an answer in a single forward pass rather than generating tokens. It accepts text and image inputs and is available on Hugging Face for local use.
• Claude Sonnet 5.5 (Anthropic): Anthropic launched Claude Sonnet 5.5 as a new model version. It is available through the Claude API and several cloud platforms.
• Claude Haiku 5.5 (Anthropic): Anthropic launched Claude Haiku 5.5, a model tuned for high-volume and latency-sensitive work. It has a 1M-token context window, supports up to 128k output tokens, and uses adaptive thinking with an effort parameter.
• Developer Knowledge API (Google): The Developer Knowledge API provides programmatic access to Google developer documentation, including documentation for Google Cloud, Firebase, and Android. It supports semantic and keyword search, document retrieval, and grounded question answering through API and developer tooling.
• Compliance API local session endpoints (Anthropic): The Compliance API local session endpoints can now provide transcripts of Claude interactions that occur in Chrome sessions (product_surface value claude_in_chrome). This capability is available in beta to Claude Enterprise organizations and integrates with existing Compliance Access Keys and scopes.
• ChatGPT Space (OpenAI): ChatGPT Space is a shared team area where teammates, ChatGPT, and a user's dot can build on shared knowledge. ChatGPT can organize the space based on provided instructions; it is available on desktop and web for eligible plans, with some page functions also available on mobile.
• Copilot code review (GitHub): Copilot code review adds a dedicated per-user settings page and more granular automatic-review controls, plus an enterprise-level default review effort that can be inherited by organizations. These changes are generally available across Copilot plans including Business and Enterprise.
• GitHub Copilot (GitHub): GitHub deprecated four models across its Copilot experiences on October 2, 2026. It suggests replacement models and says administrators may need to enable access to alternatives through model policies.
• Copilot CLI (GitHub): Copilot CLI adds whole codebase indexing (WCI) for C++ that builds a persistent symbol index across a project so the Microsoft C++ Language Server can resolve symbols and relationships without repeatedly rediscovering them. The feature is enabled by default, can increase initial build time and memory usage, and the index is reused and dynamically updated thereafter.
• Copilot for JetBrains 1.18.0 (GitHub): Copilot for JetBrains 1.18.0 is a JetBrains IDE plugin update that adds new agent features, organization-shared skills and instructions, and workflow improvements. The release includes public-preview assisted tool approvals, a re-edit message capability, Codex agent plan mode, and controls for MCP tools and the built-in MCP server.
• Playground (Google): Playground is an experimental, browser-based platform for creating, playing, and sharing custom games using text prompts, without coding experience. It supports private games, shareable links, community publishing, and multiplayer and leaderboards in select genres.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• Gemini Nano Banana 2.1 improves image generation and conversational editing, including text in images and keeping characters consistent across edits.
• Nova 2.5 Sonic adds improved reasoning, instruction following, and tool use for real-time voice agents; Strands Bidi Agents offers a way to build and connect those agents.
• Google AI Edge Foresight is an experimental Mac meeting companion that searches meeting notes and personal files using local processing, including offline.
• Mistral Large 4 is a new multimodal model for coding and other work, while qwen3.8-flash offers a 1-million-token context window for applications deployed in a US region.

Edit generated images through conversation with Gemini Nano Banana 2.1

Google’s Gemini Nano Banana 2.1 updates its image-generation model with stronger prompt following, visual quality, text rendering, and consistency for characters across multiple editing turns. It is available through the Gemini API for generating images and refining them in conversation.

Build voice agents that can follow instructions and use tools

AWS’s Nova 2.5 Sonic is a speech-to-speech model for real-time voice agents. It adds improved reasoning, instruction following, and tool-calling accuracy, with lower latency than its predecessor; developers can use it to build agents that keep track of context and use connected tools during a conversation. AWS’s Strands Bidi Agents is a platform for building production voice agents, including ones that have longer conversations or connect with other agents. It can be used with Nova 2.5 Sonic, though the record does not provide independent performance results for the combination.

Search meeting notes and personal files locally on a Mac

Google’s Google AI Edge Foresight is an experimental meeting companion for Mac that uses local AI processing to enhance notes and search across meeting transcripts and personal files. It can work offline with meeting platforms, which may suit people who want to find information in their meeting records without sending those searches to an online service.

New models target coding, image analysis, and long inputs

Mistral Large 4 is a multimodal model aimed at coding, image understanding, and professional work. Mistral also lists agent workflows and cybersecurity among its uses; these are stated capabilities, not results backed by comparisons in the record. qwen3.8-flash supports text and image analysis, function calling, and structured answers, with a 1-million-token context window. Its newly listed deployment scope is the US region in Virginia, so developers considering it should note that geographic limit.

Also on the radar

• EmbeddingGemma 2 maps text, code, images, video, and audio into a shared representation that developers can use for search, classification, or finding similar items. It was released on September 14 and entered the catalog on October 6, so it is not a new release today.
• Decisions API is a beta from OpenAI that turns text and images into typed answers; the source says it is 10 times faster than the Responses API. DiarizationLM-Gemma-4-E4B-v1 is an open model for correcting transcript turn boundaries and backchannels while keeping speaker assignments.
• Ollama v0.40.0-rc6 adds support for creating embeddings from media as well as text. Leviathan builds a searchable index from common data files and lets agents retrieve short result cards with citations.

The records

• Gemini Nano Banana 2.1 (Google): Gemini Nano Banana 2.1 is a high-efficiency image generation and conversational editing model, released as an update to Nano Banana 2. It maintains Flash-level speed and cost efficiency while improving visual quality, prompt adherence, multi-turn character consistency, text rendering, and support for wide and panoramic aspect ratios.
• EmbeddingGemma 2 (Google): EmbeddingGemma 2 is an open multimodal embedding model that maps text, code, images, video, and audio into a shared 768-dimensional vector space. It has 740 million total parameters, supports more than 100 languages, and can produce truncated embeddings at 128, 256, 512, or 768 dimensions.
• Mistral Large 4 (Mistral): Mistral Large 4 is a 1 trillion-parameter natively multimodal model with 49 billion active parameters. Mistral says it performs strongly in coding, agentic workflows, multimodal understanding, and enterprise workloads.
• qwen3.8-flash (Alibaba Qwen): qwen3.8-flash is available with a US deployment scope in the US (Virginia) region. It supports text generation, deep thinking, visual understanding, Function Calling, structured output, context caching, and a native 1M token context window.
• Nova 2.5 Sonic (AWS): Nova 2.5 Sonic is a speech-to-speech model for natural, real-time voice agents. It adds improved reasoning, instruction following, and tool-calling accuracy, with lower latency than its predecessor.
• DiarizationLM-Gemma-4-E4B-v1 (Google): DiarizationLM-Gemma-4-E4B-v1 is a 4B-parameter model fine-tuned from Gemma 4 E4B to post-process ASR and speaker-diarization outputs. The model weights are available in Hugging Face and GGUF formats, with usage instructions for Transformers and llama.cpp.
• Decisions API (OpenAI): OpenAI released the Decisions API in beta with gpt-6-luna. It turns text and images into typed answers, and the source says it is 10 times faster than the Responses API.
• Ollama v0.40.0-rc6 (Ollama): Ollama v0.40.0-rc6 adds multimodal embedding support to the MLX runner. Its /api/embed endpoint accepts media on a per-item basis through input dictionaries.
• Leviathan (elstongun): Leviathan is a single binary that builds a ranked full-text index from records in formats including JSONL, JSON, CSV/TSV, and SQLite. Agents can search the index in plain language and receive short result cards with citations; it also offers an optional MCP server.
• Strands Bidi Agents (AWS): Strands Bidi Agents is a platform for building production voice agents. It can be used with Nova 2.5 Sonic to build agents in a few lines of code, including agents with longer conversations and connections to other agents.
• Google AI Edge Foresight (Google): Google AI Edge Foresight is an experimental Mac meeting companion powered by local AI processing. It supports note enhancement, indexing, and cross-modal retrieval across meeting transcripts and personal files, and can work offline with meeting platforms.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• North 2 gives enterprises tools to build and manage agents that work with internal knowledge and tools.
• Interfaze 1 Lite is an open-weight model for processing documents, audio, and images, with features including transcription and structured extraction.
• OpenAI is inviting approved researchers and expert organizations to apply for access to a detector for text watermarks.

Build and govern workplace agents with North 2

North 2 is a platform for building and managing agents and automations that can work with an organization’s internal knowledge and tools. It includes reusable skills, shared libraries, agent memory, connectors, and governance controls, along with options for creating apps and choosing where to deploy them.

That makes it most relevant to organizations looking to create multi-step workflows and set rules around how they operate.

One open-weight model handles documents, audio, and images

Interfaze 1 Lite combines a vision-and-language reasoning core with specialist components for tasks such as speech transcription, speaker identification, visual grounding, translation, and forecasting. Developers can run it locally or use its API for multimodal tasks and structured extraction—for example, pulling information from a document or image into a consistent format.

Researchers can apply to test OpenAI’s text watermark detector

OpenAI is accepting applications for access to a detector that checks whether text contains an OpenAI watermark. Access will initially be granted case by case to approved researchers and expert organizations, who can use it to study text watermarking and its limitations.

Also on the radar

• Agens Volundr 32B Preview is an open-weights model for reasoning, coding, and tasks that use tools; Blockway says long agent sessions are a weakness, so it is better suited to shorter sessions or supervised use.
• A prerelease of Ollama automatically selects the MLX runtime for supported model architectures on Apple Silicon devices.

The records

• Interfaze 1 Lite (Interfaze AI): Interfaze 1 Lite is an open-weight mixture-of-architectures model for developer workloads, combining a vision-language reasoning core with specialist components. It supports document understanding, speech transcription and diarization, visual grounding, structured output, translation, forecasting, and safety checks.
• Agens Volundr 32B Preview (Blockway): Agens Volundr 32B Preview is an open-weights model for reasoning, coding, and agent work, with support for text, images, tool calls, and English, Chinese, and Cantonese. The preview checkpoint is available under Apache-2.0, though Blockway identifies long agentic sessions as a weakness.
• cagliostro-v3.5 (bench-labs): cagliostro-v3.5 is a 146M-parameter decoder-only language model released as downloadable float32 safetensors weights. It is an English-only base model for text completion, with a 2,048-token context length and no instruction tuning or chat template.
• Mistral 7B Instruct (Mistral): Mistral 7B Instruct is a chat model fine-tuned from Mistral 7B using publicly available instruction datasets. The source says it outperforms Llama 2 13B chat and has no moderation mechanism.
• Mistral 7B (Mistral): Mistral 7B is Mistral AI’s first 7-billion-parameter model, released under Apache 2.0. The announcement says it outperforms available open models up to 13 billion parameters on standard English and code benchmarks.
• North 2 (Cohere): North 2 is a full-stack agentic platform for enterprises, with a redesigned orchestration system for multi-step agents and automations. It includes reusable skills, shared libraries, agent memory, app creation, connectors, governance controls, and flexible deployment options.
• llama.cpp v0.5.0 (ggml-org): llama.cpp v0.5.0 is a new release of the open-source C/C++ inference library and server for running GGML-format models locally and on servers. The release adds performance optimizations, expanded model/format conversion support, server/router robustness improvements, and UI/server feature updates.
• llama.cpp b11149 (ggml-org): llama.cpp b11149 is a pre-release of the open-source C++ inference library that provides local runtimes and build artifacts for running LLMs across many platforms and backends. This release adds a -b/--backend option to test-llama-archs and ships assets and CI attestations for numerous OS/architecture/backend combinations.
• Agents (Mistral): Agents is an early version of a Mistral offering that wraps models with additional context and instructions. It is exposed through Le Chat or the API and supports creating custom behaviors and workflows using instructions and examples.
• OpenAI text watermark detector (OpenAI): OpenAI is opening applications for access to a detector that checks whether text contains an OpenAI watermark. Access will initially be granted case by case to approved researchers and expert organizations.
• ChatGPT Space (OpenAI): ChatGPT Space is a shared team area where teammates, ChatGPT, and a user's dot can build on shared knowledge. ChatGPT can organize the space based on provided instructions; it is available on desktop and web for eligible plans, with some page functions also available on mobile.
• Ollama v0.40.0-rc3 (Ollama): Ollama v0.40.0-rc3 is a pre-release of the Ollama application. It automatically uses the MLX runtime for supported model architectures on Apple Silicon devices.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• The window included new coding and decision models, releases for the llama.cpp inference library, and updates to local AI and generative media applications.

Models

Microsoft introduced FrogNano-4B-2609, a compact coding model for repository-level workflows that generates text and structured tool calls. New decision models from MaziyarPanahi and Strands Labs score or select among choices, while Aleph Alpha's Kolibri 1 BF16 is an open-weight reasoning model focused on German and English.

Developer platforms

ggml-org released llama.cpp v0.5.0 with performance optimizations, expanded model and format conversion support, and server updates. A llama.cpp pre-release added a backend option to test-llama-archs and assets for multiple operating systems, architectures, and backends.

Applications

Pre-release application Backburner distributes local Qwen3.8-27B computation between an iPhone and an Apple Silicon Mac connected over USB-C. Runway became available as a plugin in Grok Bot for image and video generation; its release note adds Ideogram 4.5 image generation in MCP workflows. GitHub made Copilot code review settings and more granular automatic-review controls generally available across Copilot plans, with API requests also supported.

The records

• FrogNano-4B-2609 (Microsoft): FrogNano-4B-2609 is a compact coding model derived from Qwen3.5-4B and post-trained with reinforcement learning on synthetic software-engineering tasks. It generates text and structured tool calls for repository-level coding workflows, with a documented context length of approximately 131K tokens.
• ModernJEV-Decide-Preview (MaziyarPanahi): ModernJEV-Decide-Preview is an experimental ModernBERT-based cross-encoder that scores choices supplied with a conversation, policy, and available tools. It returns a selected label and scores for the alternatives, but does not generate tool arguments or execute tools.
• strands-decider-2B-hobson-v19 (Strands Labs): Strands Decider is a 1.9-billion-parameter decision model for choosing among options, answering yes-or-no questions, and scoring against a rubric. It returns a confidence with each decision and is designed to run locally on GPU, Apple silicon, or CPU.
• Kolibri 1 BF16 (Aleph Alpha): Kolibri 1 BF16 is an open-weight mixture-of-experts reasoning model focused on German and English. It supports explicit reasoning, tool calling, and contexts up to 1,048,576 tokens, and is served using Aleph Alpha's inference package with vLLM.
• llama.cpp v0.5.0 (ggml-org): llama.cpp v0.5.0 is a new release of the open-source C/C++ inference library and server for running GGML-format models locally and on servers. The release adds performance optimizations, expanded model/format conversion support, server/router robustness improvements, and UI/server feature updates.
• llama.cpp b11149 (ggml-org): llama.cpp b11149 is a pre-release of the open-source C++ inference library that provides local runtimes and build artifacts for running LLMs across many platforms and backends. This release adds a -b/--backend option to test-llama-archs and ships assets and CI attestations for numerous OS/architecture/backend combinations.
• Backburner (StayLameBro): Backburner is a pre-release application that uses an iPhone connected to an Apple Silicon Mac over USB-C to help run Qwen3.8-27B locally. It provides an OpenAI-compatible server and distributes model computation and long-context attention between the Mac and phone.
• Runway MCP for Grok Bot (Runway): Runway is available as a plugin in Grok Bot. Users can sign in with a Runway account to generate images and video from a Grok Bot chat.
• Copilot code review (GitHub): Copilot code review adds a dedicated per-user settings page and more granular automatic-review controls, plus an enterprise-level default review effort that can be inherited by organizations. These changes are generally available across Copilot plans including Business and Enterprise.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• The window included new models, developer tools and applications, alongside changes to GitHub Copilot’s supported models.

Models

Google introduced Gemini 4 Argon, a frontier model with advanced reasoning and a 1-million-token output limit, designed for complex work including coding and cybersecurity defense. Cactus Compute’s Whistle runs speech recognition on a CPU and transcribes seven languages; Ai2’s open-weights AstaBrief 8B generates cited reports from research questions and retrieved literature. Cantina Security released the open-weights apex-flash-1 security model and an experimental derivative; NVIDIA’s PixelUMM handles understanding and generation across text, images and video.

Developer platforms

DeepSeek released DeepEP-Ascend, a communication library for machine-learning training and inference on Huawei Ascend NPUs. ggml-org released llama.cpp v0.5.0 and pre-release b11149, with updates to inference, model support, server features and backend testing.

Applications

Google’s Gemini app for Windows provides shortcut access and a workspace for multi-step tasks, file synthesis, and image and video generation. GitHub deprecated four models across its Copilot experiences and suggested replacements, noting that administrators may need to enable alternatives through model policies. AWS Well-Architected Agent reviews AWS infrastructure and infrastructure-as-code templates, offering prioritized recommendations and, where applicable, automation-ready fixes.

The records

• Gemini 4 Argon (Google): Gemini 4 Argon is a frontier model with advanced reasoning and a 1-million-token output limit. Google says it is designed for complex work, including coding and cybersecurity defense.
• Whistle (Cactus Compute): Whistle is a 16.9 MB speech recognition model that runs on-device on a CPU. It transcribes speech in seven languages and can also return word timestamps and speech embeddings.
• AstaBrief 8B (Ai2): AstaBrief 8B is an open-weights model that turns a research question and retrieved literature excerpts into a cited report. It is available in Asta’s Generate a report feature as Fast mode, and Ai2 is releasing the model weights and training data.
• apex-flash-1 (Cantina Security): apex-flash-1 is Cantina Security's first open-weights security model, developed in partnership with Yeta as a reinforcement-learning post-train of GLM-5.3-Flash. It is intended for focused investigations involving code reading, tool use, exploit pursuit, and verification against a running target.
• apex-flash-1-abliterated (Cantina Security): apex-flash-1-abliterated is an experimental derivative of Cantina Security’s apex-flash-1 security model, distributed as a BF16 checkpoint. It broadly modifies refusal behavior, and its image and video performance have not been evaluated.
• PixelUMM (NVIDIA): PixelUMM is an encoder-free multimodal model for understanding and generation across text, images, and video directly in pixel space. It accepts text, RGB images, and video frames, and produces text, RGB images, and video frames.
• DeepEP-Ascend (DeepSeek): DeepEP-Ascend is a communication library for machine learning training and inference on Huawei Ascend NPUs. It provides expert-parallel operations for MoE dispatch and combine, alongside additional buffer interfaces, some of which are experimental.
• llama.cpp v0.5.0 (ggml-org): llama.cpp v0.5.0 is a new release of the open-source C/C++ inference library and server for running GGML-format models locally and on servers. The release adds performance optimizations, expanded model/format conversion support, server/router robustness improvements, and UI/server feature updates.
• llama.cpp b11149 (ggml-org): llama.cpp b11149 is a pre-release of the open-source C++ inference library that provides local runtimes and build artifacts for running LLMs across many platforms and backends. This release adds a -b/--backend option to test-llama-archs and ships assets and CI attestations for numerous OS/architecture/backend combinations.
• Gemini app for Windows (Google): The Gemini app for Windows provides quick access to Gemini from a PC with the Alt + Space shortcut. Its dedicated workspace supports multi-step tasks, file synthesis, and image and video generation.
• GitHub Copilot (GitHub): GitHub deprecated four models across its Copilot experiences on October 2, 2026. It suggests replacement models and says administrators may need to enable access to alternatives through model policies.
• AWS Well-Architected Agent (AWS): AWS Well-Architected Agent analyzes AWS infrastructure and infrastructure-as-code templates against Well-Architected best practices. It provides business-goal-prioritized recommendations and, where applicable, automation-ready fixes and guided implementation materials.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• The window included new models for decisions, coding, retrieval and video, an open training framework, and GitHub Copilot updates for C++ code intelligence and local sandboxing.

Models

Perplexity introduced pplx-decider-v1-27b, a text classification model fine-tuned from Qwen3.8-27B that also accepts images and provides calibrated probabilities. Other releases include Qwen3.8-27B-pi for coding in the Pi agent harness; contextual embeddings for RAG; and Clef and Clef-Flash, multimodal models that return structured decisions. Video-related releases include a MiniMax H3 checkpoint for 2× video decoding and optional image detail enhancement, and a 19B model listed for use in a local Diffusers setup.

Developer platforms

Ai2 introduced Olmo-core 3, an open framework for developing and training large mixture-of-experts models, with a redesigned training system.

Applications

GitHub added whole codebase indexing for C++ in Copilot CLI, building a persistent symbol index that is reused and dynamically updated; the feature is enabled by default. GitHub's Copilot app added public-preview local sandboxing with per-project policies that limit filesystem, network and credential access for repository sessions.

The records

• pplx-decider-v1-27b (Perplexity): pplx-decider-v1-27b is a decision model fine-tuned from Qwen3.8-27B. It supports text classification and can also accept images for prediction.
• Qwen3.8-27B-pi (bytkim): Qwen3.8-27B-pi is a fine-tuned version of Qwen3.8-27B for coding work in the Pi agent harness. It is designed for iterative repository work, including editing files, running tools, and responding to feedback, with adjustable reasoning effort.
• MiniMax-H3-X2-Detail-v1 (speach1sdef178): A modified MiniMax H3 VAE checkpoint with two uses: 2× video decoding and optional detail enhancement of an RGB reference image. The release includes a ComfyUI custom node and workflow, and requires a separate fast decode node for the 2× workflow.
• pplx-embed-v2-context-9b-preview (Perplexity): pplx-embed-v2-context-9b-preview is a contextual embedding model for document chunks in RAG systems. It encodes document chunks together so each chunk’s embedding reflects its surrounding context, and returns one embedding per chunk.
• Clef-Flash (Cloudflare): Clef-Flash is a 9B multimodal model that maps a state and a schema of typed questions to decisions. It returns a probability for each allowed option, without free-form text generation or output parsing.
• Clef (Cloudflare): Clef is a 27B multimodal decision model that takes a state and typed questions, then returns probabilities for each allowed option. It accepts text, JSON, images, and video, and produces structured decisions rather than free-form text.
• SoL-Refiner-LTX-2.5-for-MiniMax-H3 (Efficient-Large-Model): SoL-Refiner-LTX-2.5-for-MiniMax-H3 is listed on Hugging Face as a 19B-parameter model in Safetensors format. The listing identifies a Diffusers pipeline but has no model card and is not deployed by an Inference Provider.
• Olmo-core 3 (Ai2): Olmo-core 3 is an open framework for developing and training large mixture-of-experts models. Its redesigned training system combines distributed training techniques and routing and computation optimizations, and has been benchmarked at over one trillion total parameters.
• Copilot CLI (GitHub): Copilot CLI adds whole codebase indexing (WCI) for C++ that builds a persistent symbol index across a project so the Microsoft C++ Language Server can resolve symbols and relationships without repeatedly rediscovering them. The feature is enabled by default, can increase initial build time and memory usage, and the index is reused and dynamically updated thereafter.
• Copilot app (Local sandboxing) (GitHub): Local sandboxing for the Copilot app provides per-project policies that limit filesystem, network, and credential access for local repository and working-tree sessions. The feature is available in public preview and is configurable in the app per project or for individual sessions with an in-session command.

Corrections

• GPT-6 Luna (corrected)
• GPT-6 Sol (corrected)
• Claude Opus 5.5 (corrected)

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• The window included new and retiring models, a DeepSeek port for Huawei Ascend NPUs, and updates across GitHub Copilot tools and other applications.

Models

Huihui-ai released an uncensored, proof-of-concept version of Qwen/Qwen3.8-Flash-Next created with abliteration; AgentBull introduced the open-weight Bongard-mini for scoring candidate answers or actions. Mistral's Leanstral 1.5 and OCR 4.0 are deprecated and scheduled to retire on September 30, 2026; the OCR changelog recommends OCR 4.1 instead. avbiswas released bev-decider-0.4B, a decision model that returns calibrated probabilities for text or JSON states and typed questions.

Developer platforms

DeepSeek introduced DeepGEMM Ascend, a DeepGEMM port for Huawei Ascend NPUs with compatible APIs and kernels for GEMM, MQA logits and MegaMoE.

Applications

GitHub added per-user code-review settings and more granular automatic-review controls, while local sandboxing for the Copilot app entered public preview with per-project access policies. Copilot CLI added whole-codebase indexing for C++ projects; Copilot for JetBrains 1.18.0 added agent features, shared skills and instructions, and workflow controls. AWS released a Marketplace metering agent skill for building and testing usage-based SaaS integrations. OpSafari introduced hypoarena, an offline workbench for testing hypothesis-discovery pipelines using synthetic literature.

The records

• Huihui-Qwen3.8-Flash-Next-abliterated (huihui-ai): This is an uncensored version of Qwen/Qwen3.8-Flash-Next created with abliteration. The publisher describes it as a crude, proof-of-concept implementation for removing refusals from an LLM.
• Leanstral 1.5 (Mistral): Leanstral 1.5 (labs-leanstral-1-5) is an experimental Labs model. It is deprecated and scheduled to retire on September 30, 2026.
• Bongard-mini (AgentBull): Bongard-mini is an open-weight model for machine intuition and Jev-like judgment, built on the T5Gemma 2 4B-4B encoder–decoder architecture. It takes text, structured data, or images with questions and candidate answers or actions, then returns probability distributions without generating text.
• OCR 4.0 (Mistral): OCR 4.0 (mistral-ocr-4-0) is deprecated and scheduled to retire on September 30, 2026. The changelog recommends OCR 4.1 instead and says it is available at the same price.
• bev-decider-0.4B (avbiswas): bev-decider-0.4B is a 0.4B-parameter decision model that takes text or JSON states and typed questions, returning calibrated probabilities in one forward pass without text generation. Its weights are available as a downloadable package and can be used with the bev-decider library or local server.
• DeepGEMM Ascend (DeepSeek): DeepGEMM Ascend is a port of DeepGEMM for Huawei Ascend NPUs, with an API compatible with DeepGEMM. It provides kernels for BF16, FP8, and FP4 GEMM, MQA logits, and MegaMoE.
• Copilot code review (GitHub): Copilot code review adds a dedicated per-user settings page and more granular automatic-review controls, plus an enterprise-level default review effort that can be inherited by organizations. These changes are generally available across Copilot plans including Business and Enterprise.
• Copilot CLI (GitHub): Copilot CLI adds whole codebase indexing (WCI) for C++ that builds a persistent symbol index across a project so the Microsoft C++ Language Server can resolve symbols and relationships without repeatedly rediscovering them. The feature is enabled by default, can increase initial build time and memory usage, and the index is reused and dynamically updated thereafter.
• Copilot app (Local sandboxing) (GitHub): Local sandboxing for the Copilot app provides per-project policies that limit filesystem, network, and credential access for local repository and working-tree sessions. The feature is available in public preview and is configurable in the app per project or for individual sessions with an in-session command.
• Copilot for JetBrains 1.18.0 (GitHub): Copilot for JetBrains 1.18.0 is a JetBrains IDE plugin update that adds new agent features, organization-shared skills and instructions, and workflow improvements. The release includes public-preview assisted tool approvals, a re-edit message capability, Codex agent plan mode, and controls for MCP tools and the built-in MCP server.
• AWS Marketplace metering agent skill (AWS): The AWS Marketplace metering agent skill guides sellers through building, deploying, and validating usage-based SaaS metering integrations. It can recommend a metering API, generate integration code and a customized CloudFormation stack, and run an end-to-end test against AWS Marketplace.
• hypoarena (OpSafari): hypoarena is an offline scientific hypothesis-discovery workbench that provides components for building and testing generate–debate–evolve pipelines. Its demo runs on synthetic literature and explicitly makes no claim about real scientific discovery capability.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• The window included new models for reasoning, motion generation, tabular prediction and video, alongside developer tools and applications for collaboration, agents and data engineering.

Models

OpenAI introduced GPT-6.1 Sol for coding, professional work, computer use and multi-step workflows. New releases also include PostHog’s Jeeves reasoning classifier, UniMate motion generation model, NVIDIA’s Kumo Tabular, Jeff-Qwen3.5-0.8B, Gemini Omni Flash and Meta’s Muse Video preview.

Developer platforms

OpenAI’s Plugin extensions let developers build ChatGPT experiences with sidebar homes, interactive panels and viewers for supported file types.

Applications

OpenAI introduced ChatGPT Space for shared team knowledge and Dots, always-on agents that connect to apps through plugins. The open-source iCode platform provides terminal-based agent and workflow tools, while Microsoft previewed an autonomous agent for long-running data engineering tasks.

The records

• Jeeves (PostHog): Jeeves is a 9B reasoning classifier based on Qwen3.5-9B, with a pointer head and a block-4 diffusion drafter. Released weights, training code, and a Jev-compatible API are provided.
• UniMate unimate_uniml3d_f60_v2 (Linzhan Mou): UniMate unimate_uniml3d_f60_v2 is a text-conditioned flow-matching model that generates motion for skeletons with 5 to 70 joints. It produces 60-frame clips and was trained on animal, humanoid, and rigged-object motion data.
• GPT-6.1 Sol (OpenAI): GPT-6.1 Sol is a model for coding, professional work, computer use, and multi-step workflows. OpenAI says it improves on GPT-6 Sol and approaches GPT-6 Astra on several evaluations at substantially lower cost.
• Kumo Tabular (NVIDIA): Kumo Tabular is NVIDIA's pretrained tabular foundation model for classification and regression. Its model weights are released under the OpenMDW 1.1 license, and inference is available through the structured-data-models package.
• Jeff-Qwen3.5-0.8B (firelex): Jeff-Qwen3.5-0.8B is a fine-tuned model for zero-shot classification that returns calibrated probabilities for described options in one forward pass. Its weights are available on Hugging Face, and the repository provides instructions to run it locally.
• Gemini Omni Flash (Google): The release note directs users to the Gemini Omni Flash model page and the omni guide to get started. It does not describe the model’s capabilities, pricing, or deployment details.
• Muse Video (Meta): Muse Video is an early preview video-generation model built on the same pretraining base as Muse Image, delivering visual fidelity and native audio support with competitive prompt adherence and temporal consistency. It is being previewed for creators and integration into Meta AI with further work planned on audio-video synchronization and fast-motion physical accuracy.
• Plugin extensions (OpenAI): Plugin extensions let developers create experiences within ChatGPT, including a sidebar home and interactive panels alongside conversations. Developers can also create viewers for file types their products support, and the extensions are available to all plans.
• ChatGPT Space (OpenAI): ChatGPT Space is a shared team area where teammates, ChatGPT, and a user's dot can build on shared knowledge. ChatGPT can organize the space based on provided instructions; it is available on desktop and web for eligible plans, with some page functions also available on mobile.
• Dots (OpenAI): Dots are always-on AI agents powered by GPT-6 Astra that work on their own cloud computers and connect to apps through plugins. Users can message or call them in ChatGPT and reach them through Slack and Teams, while setting permissions and reviewing their work.
• iCode (openJiuwen-ai): iCode is an open-source, offline development platform and agent/workflow toolkit with a terminal interface. It lets users configure agents and model profiles, run workflows, and review changes made to files.
• Fabric data engineering agent (Microsoft): The Fabric data engineering agent is a preview offering for complex, long-running data engineering work. It can plan and execute tasks autonomously within engineer-defined outcomes and guardrails, validate results, and refine its approach.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• The window also brought new decision, agentic and multimodal models, alongside a llama.cpp release and pre-release. A separate correction was listed.

Models

ElevenLabs introduced v4 Turbo, a low-latency text-to-speech model, and v4, which offers controls for directing speech delivery. Both support more than 90 languages. New releases include Julia 1 and Lumma-Fev-4B, decision models that score supplied options, and Jev-Omni, an open-weight multimodal decision classifier. Anthropic launched Claude Sonnet 5.5, available through the Claude API and several cloud platforms. Aratako released Irodori-TTS-v4-Large, a Japanese speech model with voice cloning and design features. Hcompany introduced Holo4-27B for software interaction, while Xiaomi MiMo released MiMo-V2.6-Pro-MOPD for multimodal tasks and agent workflows. IQuest-Q1 targets agentic coding, reasoning and tool use.

Developer platforms

ggml-org released llama.cpp v0.5.0 with performance and conversion updates, plus improvements to server and router robustness. The llama.cpp b11149 pre-release adds a backend option to test-llama-archs and includes assets and CI attestations across operating systems, architectures and backends.

The records

• Eleven v4 Turbo (ElevenLabs): Eleven v4 Turbo is a low-latency text-to-speech model that combines expressive speech generation with faster response times. It supports more than 90 languages and is designed to work with ElevenLabs’ conversational agents platform.
• Eleven v4 (ElevenLabs): Eleven v4 is a text-to-speech model designed to interpret tone, pacing, emotion, character, and context. It supports more than 90 languages and includes controls for directing speech delivery with natural-language prompts and audio tags.
• Julia 1 (Supersonic Labs): Julia 1 is a 144.3M-parameter decision model based on JHU CLSP's mmBERT-small, released with checkpoint weights and inference code. It scores supplied answer options for classification, routing, ordered scores, and Boolean decisions; it is not a text-generation model.
• Claude Sonnet 5.5 (Anthropic): Anthropic launched Claude Sonnet 5.5 as a new model version. It is available through the Claude API and several cloud platforms.
• Irodori-TTS-v4-Large (Aratako): Irodori-TTS-v4-Large is a Japanese text-to-speech model that supports voice cloning and voice design using text, reference speech, and caption text. It also supports emoji-based style control and integrates invisible audio watermarking.
• Holo4-27B (Hcompany): Holo4-27B is a 27B dense agentic model that interacts with software through GUIs, code, MCP and APIs. It is available through the H Models API and as downloadable weights.
• Lumma-Fev-4B (FrontiersMind): Lumma-Fev-4B reads a document or other state together with typed questions and returns probability distributions in a single forward pass. It does not generate text, and its weights are available under Apache-2.0.
• Jev-Omni (akhilaaa3): Jev-Omni is an open-weight multimodal decision classifier built on Gemma 4 12B IT and fine-tuned on 30,000 questions. It returns probabilities for supplied options rather than generating an explanation.
• MiMo-V2.6-Pro-MOPD (Xiaomi MiMo): MiMo-V2.6-Pro-MOPD is a multimodal sparse MoE model with 1.02T total parameters, 42B activated parameters, and a context length of up to 1M tokens. Its MOPD training combines domain-specialized teachers and is intended to reduce repeated tool calls in agentic settings.
• IQuest-Q1 (IQuest): IQuest-Q1 is a text-only Mixture-of-Experts model developed for agentic coding, reasoning, and multi-step tool use. It has approximately 320 billion total parameters, with an estimated 15 billion activated per token, and a context length of 524,288 tokens.
• llama.cpp v0.5.0 (ggml-org): llama.cpp v0.5.0 is a new release of the open-source C/C++ inference library and server for running GGML-format models locally and on servers. The release adds performance optimizations, expanded model/format conversion support, server/router robustness improvements, and UI/server feature updates.
• llama.cpp b11149 (ggml-org): llama.cpp b11149 is a pre-release of the open-source C++ inference library that provides local runtimes and build artifacts for running LLMs across many platforms and backends. This release adds a -b/--backend option to test-llama-archs and ships assets and CI attestations for numerous OS/architecture/backend combinations.

Corrections

• Ray-Ban Meta Audio (corrected)

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.

TL;DR

• The window included two 309B-parameter models with 1M-token context windows, alongside models for structured decisions and CT analysis.

Models

Xiaomi MiMo introduced MiMo-V2.6-Flash-MOPD, a multimodal sparse MoE model with 309B total parameters, 15B activated parameters and a 1M-token context length. It is an upgrade intended to mitigate repeated tool calls. NaiveAI introduced Naive-N0.5-Flash, an open-weight 309B mixture-of-experts model with 15.5B active parameters and a native 1M-token context window, designed for coding and AI R&D. InternLM introduced Intern-Decision-0.8B, which scores candidate answers across schema-defined questions in one pass and returns calibrated probabilities in typed JSON. NVIDIA introduced NV-Reason-CT for chest and abdominal CT analysis, including report generation and question answering.

The records

• MiMo-V2.6-Flash-MOPD (Xiaomi MiMo): MiMo-V2.6-Flash-MOPD is a multimodal sparse MoE model with 309B total parameters and 15B activated parameters. It has a 1M-token context length and is an MOPD upgrade of the MiMo-V2.6-Flash-RL checkpoint that mitigates repeated tool calls.
• Naive-N0.5-Flash (NaiveAI): Naive-N0.5-Flash is an open-weight 309B mixture-of-experts model with 15.5B active parameters, designed for coding and AI R&D. It has a native 1M-token context window and uses hybrid Sliding-Window Attention and DeepSeek Sparse Attention.
• Intern-Decision-0.8B (InternLM): Intern-Decision-0.8B is a multimodal structured decision model fine-tuned from Qwen3.5-0.8B. It scores candidate answers for multiple schema-defined questions in one forward pass and returns calibrated probabilities in typed JSON.
• NV-Reason-CT (NVIDIA): NV-Reason-CT is a 3D vision-language model for chest and abdominal CT analysis, combining a native 3D vision encoder with a Qwen3.5-4B language model. It supports report generation, question answering, and reasoning, and is released for research and educational use.

CC BY 4.0, attributed to wowsuchai (https://wowsuchai.withglint.app). Linked source content keeps its original rights.