Yesterday, Liquid AI released LFM2.5-VL-3B, a 3.1B-parameter vision-language model designed to run on mobile, web, and desktop devices. The system reads digital screens, locates objects with specific coordinates, parses documents, and executes tools based on text or image input. The model scores an average of 69.4 across 28 vision benchmarks, matching the performance of the 4.7B InternVL-3.5-4B model and trailing Qwen3.5-4B by 0.7 points. It lacks reasoning capabilities, answering queries directly to maintain low latency. The model occupies roughly 3 GB of memory and decodes at 228 tokens per second on an Apple M5 Max.
Is it deployable?
The checkpoint is available in four formats: native, GGUF, ONNX, and MLX. Day-one runtimes include llama.cpp, MLX, vLLM, SGLang, and ONNX. The 3 GB memory footprint allows deployment on standard consumer hardware.
- Licensing: The LFM Open License v1.0 is based on Apache-2.0 with one restriction. Commercial use is free for companies with annual revenue under $10 million USD. Enterprises exceeding this threshold must negotiate a commercial license with Liquid AI. Research, education, and non-profit use have no revenue cap.
- Target industries: Consumer electronics, automotive, industrial and robotics, financial services, healthcare, and e-commerce. The model also serves QA and RPA vendors automating graphical user interfaces.
- Use cases: On-device screen agents, GUI test automation, PDF-to-structured-text conversion 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.
What changed?
LFM2.5-VL-3B extends the previous LFM2-VL-3B model across four areas.
- 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 leading at 84.1.
- Function calling: This capability is new to the VL line. ToolSandbox scores moved from 26.4 to 59.5. BFCL v4 improved from 20.5 to 32.5. Tool calls appear as Pythonic functions between
<|tool_call_start|>and<|tool_call_end|>tokens. - Grounding: RefCOCO-avg precision@1 rose from 57.1 to 87.9, a 30-point gain driven by scaled synthetic grounding data.
- Multi-image input: BLINK improved 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 uses 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 expanded to 128K by extending the existing tokenizer in place, which improves coverage for non-Latin scripts. Vision pre-training scaled 4× in tokens with curated and synthetic caption, OCR, grounding, and instruction-following data.
Post-training involves supervised fine-tuning with knowledge distillation from a larger teacher and Antidoom training, followed by multi-reward reinforcement learning. The model is non-reasoning, answering directly to support its latency profile.
Benchmarks
Liquid AI evaluated the model 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 include RealWorldQA at 73.1 against InternVL-3.5-4B (67.7), TextVQA at 84.3 against Qwen3.5-4B (81.2), MMStar at 63.3, MathVista-mini at 68.5, ChartQA at 81.3, DocVQA at 91.1, and OCRBench v1 at 84.2. CountBenchQA regressed to 87.3 from 92.2 in the prior release.
On text-only evaluation, IFEval reached 82.3, up from 72.9. Gemma-4-E4B still leads there at 87.9.
What it means
Developers can now ship on-device agents that interact with GUIs without sending data to the cloud. The tool-calling feature allows the model to trigger external functions directly from screen content, which is useful for automating repetitive tasks on a phone or tablet. The licensing model permits commercial use for smaller businesses and startups without upfront fees, though larger enterprises must pay for a license.
