Liquid AI launches 3-billion-parameter vision-language model for on-device use

Liquid AI released LFM2.5-VL-3B, a 3.1 billion-parameter vision-language model built from the ground up to run on-device. The model can read digital screens on mobile, web and desktop, locate objects in coordinates, extract information from documents and diagrams, and invoke function calling directly from text or images. Its average score of 69.4 points across 28 vision benchmarks puts it on par with InternVL-3.5-4B and 0.7 points behind Qwen3.5-4B, both of which are 1.6 billion parameters larger.

In isolated tests the model scores 73.1 on RealWorldQA versus 67.7 for InternVL, 84.3 on TextVQA versus 81.2 for Qwen, 91.1 on DocVQA and 84.2 on OCRBench v1. On the UI front it achieves an 80.7 average on ScreenSpot-v2 (desktop 78.7, mobile 81.2, web 82.2), outpacing Gemma-4-E4B at 51.2 and Qwen3.5-4B at 78.5, and falling only behind InternVL-3.5-4B at 84.1. The sole recorded weakness is a regression on CountBenchQA, which dropped from 92.2 in the previous version to 87.3 now. In text-only evaluation IFEval jumped from 72.9 to 82.3, though Gemma-4-E4B still leads at 87.9.

The language backbone is LFM2.5-2.6B, while the vision tower is a customized SigLIP2 NaFlex encoder of 400 million parameters. NaFlex handles original resolution by splitting large images into non-overlapping 512×512 patches plus a down-sampled thumbnail of the whole image. The context length is 32,768 tokens with support for 16 languages. Pre-training consumed roughly 34 trillion tokens; the vocabulary was doubled to 128 thousand entries by expanding the existing tokenizer in place, improving coverage of non-Latin scripts. Vision training was quadrupled in token count with synthetic and curated caption, OCR, grounding and instruction data. Post-training included SFT distilled from a larger teacher model, Antidoom training, and multi-reward reinforcement learning. The design choice to make the model non-reasoning—answering directly without a chain-of-thought—stems from the low-latency profile.

The checkpoint is available in four formats at launch: native, GGUF, ONNX and MLX, with immediate support for llama.cpp, MLX, vLLM, SGLang and ONNX. The model occupies about 3 GB of memory and decodes 228 tokens per second on an Apple M5 Max. The LFM Open License v1.0 is Apache-2.0-based with a single amendment: free commercial use ends when the company’s annual revenue exceeds $10 million (≈ 37 million shekels). Indie developers, startups and SMBs below that threshold may embed the model without payment; larger organizations must negotiate a commercial license with Liquid AI. Research, educational and non-profit use remains unrestricted.

Declared targets span consumer electronics, automotive, industrial and robotics, finance, health, e-commerce and QA/RPA providers that automate user interfaces. Use cases include on-device screen agents, GUI test automation, PDF-to-structured-text conversion with layout tags, invoice and receipt OCR, real-time object detection in vehicles, offline translation of menus and road signs, and multi-image comparison. Function calling is exposed via a Python-style interface between dedicated tokens, and grounding accuracy leapt from 57.1 to 87.9 on RefCOCO, a 30-point gain attributed to large-scale synthetic data. In short, a compact, fast model with a small-developer-friendly license that narrows the gap with larger competitors on practical vision tasks.