Yesterday, Liquid AI released LFM2.5-VL-3B. It is a 3.1B-parameter vision-language model built for on-device deployment. The model reads digital screens across mobile, web, and desktop. It grounds objects to coordinates, parses documents and charts, and calls tools from text or image input. Liquid AI reports an average of 69.4 across 28 vision benchmarks. That matches InternVL-3.5-4B and sits 0.7 points behind Qwen3.5-4B, both 4.7B models. The model is non-reasoning, so it answers directly and keeps latency low. It fits in roughly 3 GB of memory and decodes 228 tokens/s on an Apple M5 Max.
Is it deployable?
Yes, the checkpoint ships in four formats: native, GGUF, ONNX, and MLX. Day-one runtimes include llama.cpp, MLX, vLLM, SGLang, and ONNX. It fits in roughly 3 GB of memory.
- Which company levels: The LFM Open License v1.0 is Apache-2.0-based with one change: free commercial use ends once a company’s annual revenue reaches $10M USD. So indie developers, startups, and SMBs under that line can ship commercially at no cost. Enterprises above it must negotiate a commercial license with Liquid AI. Research, education, and non-profit use carry no revenue limit.
- Industries: Consumer electronics, automotive, industrial and robotics, financial services, healthcare, and e-commerce. Also QA and RPA vendors that automate GUIs.
- Applications: On-device screen agents, GUI test automation, PDF-to-structured-text with layout labels, invoice and receipt OCR, near-real-time object detection in vehicles, offline translation of menus and road signs, and multi-image comparison.
So, What is new?
LFM2.5-VL-3B extends LFM2-VL-3B along four axes.
- Screen and UI understanding: The model averages 80.7 on ScreenSpot-v2 across desktop (78.7), mobile (81.2), and web (82.2). Liquid AI reports Gemma-4-E4B at 51.2 and Qwen3.5-4B at 78.5, with the larger InternVL-3.5-4B ahead at 84.1.
- Function calling: This is new to the VL line. ToolSandbox moves from 26.4 to 59.5. BFCL v4 moves from 20.5 to 32.5. Tool calls are emitted as Pythonic calls between
<|tool_call_start|>and<|tool_call_end|>tokens. - Grounding: RefCOCO-avg precision@1 rises from 57.1 to 87.9, a 30-point gain driven by scaled synthetic grounding data.
- Multi-image input: BLINK improves from 50.2 to 61.5, and MuirBench from 34.9 to 58.3.
Architecture and training
The language backbone is LFM2.5-2.6B. The vision tower is a SigLIP2 NaFlex shape-optimized 400M encoder. NaFlex handles native resolution by splitting large images into non-overlapping 512×512 patches plus a resized whole-image thumbnail. Context length is 32,768 tokens, and 16 languages are supported.
Pre-training used approximately 34T tokens. Vocabulary was doubled to 128K by extending the existing tokenizer in place, which improves non-Latin script coverage. Vision pre-training was scaled 4× in tokens with curated and synthetic caption, OCR, grounding, and instruction-following data.
Post-training is SFT with knowledge distillation from a larger teacher and Antidoom training, followed by multi-reward reinforcement learning.
The model is non-reasoning. It answers directly, which is the design choice behind its latency profile.
Benchmarks
Liquid AI evaluated across 28 vision benchmarks using vLLM 0.26.0 in non-reasoning mode. LFM2.5-VL-3B averages 69.4, matching InternVL-3.5-4B (69.4) and landing 0.7 points behind Qwen3.5-4B (70.1). Both comparison models are 4.7B parameters.
Notable individual results: RealWorldQA 73.1 against InternVL-3.5-4B at 67.7, TextVQA 84.3 against Qwen3.5-4B at 81.2, MMStar 63.3, MathVista-mini 68.5, ChartQA 81.3, DocVQA 91.1, and OCRBench v1 84.2. CountBenchQA regressed to 87.3 from 92.2 in the prior release.
On text-only evaluation, IFEval reaches 82.3, up from 72.9. Gemma-4-E4B still leads there at 87.9.
