LFM2.5-VL-3B

LFM2.5-VL-3B is Liquid AI's open-weight vision-language model (~3.12B parameters; HF safetensors total 3,123,483,888) for on-device and edge image-text→text. Non-reasoning: answers directly for low latency. Builds on the LFM2.5-2.6B language backbone with a SigLIP2 400M NaFlex vision encoder; improves screen/UI understanding, grounding, function calling, and multi-image input over LFM2-VL-3B. Pre-trained on ~34T tokens; 128K vocabulary; 32,768-token context. Day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX. License: LFM Open License v1.0 (lfm1.0). Catalog benches include MMStar, MME (Liquid 0–100 normalize of 0–2800), RealWorldQA, SimpleVQA, MMBench-V1.1, MM-IF-Eval, MathVista-Mini, MMMU-Pro, MMMU (val), ChartQA, DocVQA, OCRBench, OCRBench-V2 (en), TextVQA, RefCOCO-avg, BLINK, MuirBench, HallusionBench, POPE, IFEval, IFBench, Multi-IF, and BFCL-v4. No catalog id for ScreenSpot-v2 (80.7 avg), CountBenchQA, SEED-Bench (image), MMMB, Multilingual MMBench, LogicVista, InfographicVQA, or ToolSandbox — recorded here only. Sources: https://www.liquid.ai/blog/lfm2-5-vl-3b · https://huggingface.co/LiquidAI/LFM2.5-VL-3B. No hosted provider pricing row (self-host open weights; OpenRouter has only liquid/lfm-2.5-2.6b:free for the text sibling).