Liquid AI's 3B model matches 4.7B rivals on vision benchmarks

Liquid AI just dropped LFM2.5-VL-3B, a 3.1B-parameter vision-language model that punches well above its weight class. Across 28 vision benchmarks, it averages 69.4 — dead even with InternVL-3.5-4B and just 0.7 points shy of Qwen3.5-4B, both of which carry roughly 50% more parameters. The model is deliberately non-reasoning, which means it answers directly instead of thinking out loud. That design choice keeps latency low enough for real-world deployment: roughly 3 GB of memory and 228 tokens per second on an Apple M5 Max, or 20 tokens per second on a Galaxy S26 Ultra.

The headline numbers are impressive, but the real story is the 30-point jump in grounding. RefCOCO-avg precision@1 went from 57.1 in the prior release to 87.9, a leap Liquid AI attributes to scaled synthetic grounding data. Screen understanding also got a serious upgrade — ScreenSpot-v2 now averages 80.7 across desktop, mobile, and web, beating Qwen3.5-4B’s 78.5 and leaving Gemma-4-E4B in the dust at 51.2. InternVL-3.5-4B still edges it out at 84.1, but at nearly double the parameter count.

Function calling is brand new to the VL line, and the gains are substantial if uneven. ToolSandbox jumped from 26.4 to 59.5, and BFCL v4 climbed from 20.5 to 32.5. Tool calls come wrapped in dedicated tokens as Pythonic calls, which should make integration cleaner for developers building agentic workflows. The architecture pairs LFM2.5-2.6B as the language backbone with a SigLIP2 NaFlex 400M vision encoder that splits large images into 512×512 patches plus a resized thumbnail — a practical approach for handling native resolution without blowing up context. Speaking of context, it supports 32,768 tokens and 16 languages.

One wrinkle worth flagging: CountBenchQA regressed from 92.2 to 87.3. Regression on a counting benchmark while everything else improves suggests the synthetic data scaling may have traded some precision in specific visual reasoning tasks for broader gains. It’s a minor blemish, but worth watching if your use case involves precise object counting. On the text side, IFEval climbed to 82.3 from 72.9, though Gemma-4-E4B still leads at 87.9.

The licensing is where things get interesting. LFM Open License v1.0 is Apache-2.0-based with one twist: free commercial use ends once your company hits $10M in annual revenue. That’s a clean line for indie devs and startups — ship commercially at no cost, no legal ambiguity. Enterprises above the threshold need to negotiate with Liquid AI directly. For QA and RPA vendors automating GUI testing, on-device screen agents, or PDF-to-structured-text pipelines, this model is genuinely deployable day one across llama.cpp, MLX, vLLM, SGLang, and ONNX. The 4× scale-up in vision pre-training tokens is the quiet engine behind most of these gains, and it shows.