Liquid AI releases 3.1B-parameter vision-language model for on-device use
Liquid AI has released LFM2.5-VL-3B, a 3.1B-parameter vision-language model designed to run on devices. It can interpret mobile, web and desktop screens, identify objects by coordinates, process documents and charts, and call tools using text or image input. The non-reasoning model is intended to answer directly with low latency, using roughly 3 GB of memory. On an Apple M5 Max, it decodes 228 tokens per second.
Across 28 vision benchmarks, the model averages 69.4, matching InternVL-3.5-4B and trailing Qwen3.5-4B by 0.7 points. ScreenSpot-v2 averages 80.7, while RefCOCO-avg precision@1 increased from 57.1 in the previous release to 87.9. ToolSandbox rose from 26.4 to 59.5, and BFCL v4 from 20.5 to 32.5. BLINK and MuirBench also improved to 61.5 and 58.3, respectively.
The checkpoint is available in native, GGUF, ONNX and MLX formats, with support for llama.cpp, MLX, vLLM, SGLang and ONNX runtimes. Its LFM Open License v1.0 permits commercial use without charge below $10M in annual revenue; larger companies must negotiate a license. Research, education and nonprofit use have no revenue limit.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.