Liquid AI introduces LFM2.5-VL-3B for fast edge vision tasks
Liquid AI has introduced LFM2.5-VL-3B, a 3.1B-parameter vision-language model designed to run on local hardware. The model handles documents and screens, identifies objects through natural-language queries, processes multiple images, and supports function calling. It is designed to answer directly rather than perform extended reasoning, which the company says helps keep responses fast for real-time, on-device applications.
LFM2.5-VL-3B combines a SigLIP2 400M NaFlex vision encoder with the pre-trained backbone of the LFM2.5-2.6B text model. Training used about 34T tokens and four times more vision data than previous releases. The vocabulary was expanded to 128K tokens to support non-Latin scripts. Post-training included supervised fine-tuning with knowledge distillation and Antidoom training, followed by multi-reward reinforcement learning.
In Liquid AI’s evaluations, the model led its size class on several real-world image tasks and achieved an overall vision benchmark average of 69.4. It supports llama.cpp, MLX, vLLM, SGLang, and ONNX. The model uses about 3 GB of memory, reaches 228 tokens/s on an M5 Max and 20 tokens/s on a Galaxy S26 Ultra, and is available on Hugging Face with a browser-based WebGPU demo.
Written by Sakhanda Wire AI from reporting by Hugging Face. Facts, quotes and full context are in the original article below.