LFM2.5-VL-3B

Use it for image understanding, object detection, OCR with layout annotation, document processing, or on-device translation.

Availability

Checkpoint and quantized exports are linked on Hugging Face; browser demos are available.

Pricing model

Not established

Deployment

Designed for on-device deployment; ONNX exports support cloud, edge, and mobile.

Privacy

WebGPU demo runs entirely on-device

Licence

Open-weight; download, fine-tune, and deploy without restrictions

• LFM2.5-VL-3B is available on Hugging Face and Liquid AI Playground. ^[1]

  The model is available today on Hugging Face and our Playground .

• The model improves screen understanding, grounding, function calling, and multi-image input over LFM2-VL-3B. ^[1]

  LFM2.5-VL-3B builds on our previous LFM2-VL-3B , with significant improvements in screen understanding, grounding, function calling, and multi-image input.

• The model is non-reasoning and answers directly to keep latency low. ^[1]

  LFM2.5-VL-3B is a non-reasoning model that answers directly, keeping latency low for real-time and on-device applications.

• The WebGPU demo runs entirely on-device with no setup required. ^[1]

  The model runs entirely on-device with no setup required.

• The model is open-weight and can be downloaded, fine-tuned, and deployed without restrictions. ^[1]
• The model processes both text and images. ^[2]

  LFM2.5-VL-3B can process both text and images, and uses the LFM2.5-2.6B language model as its backbone, combined with a SigLIP2 NaFlex vision encoder.

• It supports object detection and grounding with natural-language queries. ^{[1] [2]}

  * **Better grounding**: Improved grounding and object detection with natural language queries.

• It supports full-page OCR with layout annotation. ^[2]

  * **Better OCR**: Full page OCR with layout annotation.

• The original checkpoint and GGUF, ONNX, and MLX exports are linked for download.
• The model family is designed for on-device deployment.
• Liquid AI recommends it for single-turn, high-throughput, low-latency tasks and not for long-context, reasoning-intensive tasks. ^[2]

  We recommend using it for single-turn, high-throughput, low-latency tasks; for example, for near-realtime object detection in automotive applications, batch processing scanned documents with OCR with layout information for turning PDFs into searchable text, or for on-device translation of menus and road signs into your native language.