LFM3-VL-2B-VLD Multipass

A person can use it to make decisions from browser workflow screenshots, including verifying options and producing an option-code readout.

Availability

Available as a Hugging Face repository

Pricing model

Not established

Deployment

Companion ONNX bundle for WebGPU and GGUF exports for llama.cpp

Privacy

Not established

Licence

other

• The model is published in the webbrain-one/lfm3-vld-2b-multipass Hugging Face repository.
• The model uses a multipass VL-D protocol with a Yes/No verification step per option and an option-code readout. ^[1]

  `LiquidAI/LFM3-VL-2B` adapted with WebBrain's multipass **VL-D** (`verify` Yes/No per option + `readout` option code) decision protocol, distilled from `LiquidAI/d1-2B` on WebBrain multimodal browser workflow screenshots — **without full VL/Compass fine-tuning** (vision tower frozen, language backbone adapted via LoRA and merged to BF16).

• The model was distilled from LiquidAI/d1-2B on multimodal browser workflow screenshots. ^[1]

  `LiquidAI/LFM3-VL-2B` adapted with WebBrain's multipass **VL-D** (`verify` Yes/No per option + `readout` option code) decision protocol, distilled from `LiquidAI/d1-2B` on WebBrain multimodal browser workflow screenshots — **without full VL/Compass fine-tuning** (vision tower frozen, language backbone adapted via LoRA and merged to BF16).

• On the d1 Balanced Browser Decision Benchmark final_test, probe/probe-cs scored 126/126, or 100.0%, correct. ^[1]

  ## Results on d1 Balanced Browser Decision Benchmark (`final_test`, 126 Categorical Questions)

• A companion ONNX repository includes a WebGPU bundle. ^[1]
• Companion GGUF exports are provided for llama.cpp. ^[1]

  - [`webbrain-one/lfm3-vld-2b-multipass-gguf`](https://huggingface.co/webbrain-one/lfm3-vl-2b-vld-gguf): GGUF exports (`BF16`, `Q8_0`, `Q4_K_M`, and `mmproj-F16`) for `llama.cpp` (`/v1/chat/completions` + `/v1/systemone`).