LFM2.5-VL-3B for Better and Faster Vision Capabilities for the Edge

Hugging Face · 2 months ago · 22 ● 2 sources

Liquid AI released LFM2.5-VL-3B, a vision-language model that runs on users’ own hardware and supports screen understanding, object grounding, multi-image inputs, and function calling for real-time on-device use. It was pre-trained on about 34T tokens and uses a 128K vocabulary extension for non-Latin scripts. The release adds faster, edge-friendly inference (including ~228 tokens/s on an M5 Max and ~3 GB memory fit) alongside improved UI/screen comprehension, grounding, and tool/function calling performance versus its prior models in this series.