Liquid AI Releases LFM2.5-VL-3B: A 3B Vision-Language Model That Reads Screens, Grounds Objects, and Calls Tools On-Device
| Source: MarkTechPost
Tags: Liquid AI, LFM2.5-VL-3B, vision-language model, on-device AI, function calling, MLX, GGUF, screen understanding
Liquid AI's LFM2.5-VL-3B is a 3.1B-parameter on-device vision model that achieves 69.4 across 28 vision benchmarks — matching 4.7B models — while running at 228 tokens/s on an Apple M5 Max in 3GB of memory, with function calling and a 30-point grounding accuracy gain.
Details
Liquid AI released LFM2.5-VL-3B, a 3.1B-parameter vision-language model optimized for on-device deployment. Despite being smaller than comparable models, it scores 69.4 on average across 28 vision benchmarks — matching InternVL-3.5-4B and sitting 0.7 points behind Qwen3.5-4B, both 4.7B models. On an Apple M5 Max, it decodes 228 tokens per second and fits in approximately 3 GB of memory. Four key improvements over its predecessor: grounding precision jumped from 57.1 to 87.9 on RefCOCO-avg (a 30-point gain from scaled synthetic training data), function calling was added for the first time in this model line (ToolSandbox rising from 26.4 to 59.5), multi-image understanding improved substantially (MuirBench from 34.9 to 58.3), and screen-understanding averages 80.7 on ScreenSpot-v2. The model ships in four formats (native, GGUF, ONNX, MLX) with day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX. The license is Apache 2.0-based with one threshold: free commercial use for companies under $10M annual revenue; enterprises above that must negotiate a paid license. For developers building GUI automation, screen agents, or document OCR pipelines, this is the strongest on-device option at this parameter count.