Liquid AI’s On-Device Vision-Language Model Hits 69.4 on 28 Benchmarks
A 3.1-billion-parameter model that fits in 3 GB of RAM and runs at 228 tokens per second on an Apple M5 Max is now available for on-device deployment—no internet required. Liquid AI’s LFM2.5-VL-3B reads digital screens, grounds objects to pixel coordinates, parses documents and charts, and even calls tools from text or image input, all within the device’s memory budget. On 28 vision-language benchmarks, it averages 69.4, matching the larger InternVL-3.5-4B and trailing Qwen3.5-4B by just 0.7 points despite being 1.6 billion parameters smaller.
How it works
Under the hood, LFM2.5-VL-3B pairs a 2.6B-parameter language backbone with a 400M-parameter SigLIP2 vision encoder optimized for native resolution. NaFlex splits large images into 512×512 patches plus a resized thumbnail, letting the model handle native resolutions without rescaling artifacts. Context length reaches 32,768 tokens across 16 languages, and the tokenizer now covers non-Latin scripts thanks to a doubled vocabulary of 128,000 tokens. Training relied on roughly 34 trillion tokens, with vision pre-training quadrupled in data volume and post-training combining supervised fine-tuning, knowledge distillation, and multi-reward reinforcement learning.
Licensing and reach
Liquid AI ships the checkpoint in four formats—native, GGUF, ONNX, and MLX—with day-one runtimes including llama.cpp, MLX, vLLM, SGLang, and ONNX. The LFM Open License v1.0 follows Apache 2.0 but caps free commercial use at $10M annual revenue; research, education, and non-profits face no revenue limit. Target sectors span consumer electronics, automotive, industrial robotics, financial services, healthcare, and e-commerce, as well as QA and RPA vendors automating GUIs.
What’s new
Compared with its predecessor, LFM2-VL-3B, the new model gains screen and UI understanding (+80.7 on ScreenSpot-v2), tool-calling capability (ToolSandbox jumps from 26.4 to 59.5, BFCL v4 from 20.5 to 32.5), stronger object grounding (RefCOCO precision@1 rises from 57.1 to 87.9), and multi-image comprehension (BLINK +11.3, MuirBench +23.4).
Why it matters
For developers targeting edge devices, LFM2.5-VL-3B lowers latency and eliminates cloud dependency while preserving competitive accuracy. The licensing model keeps costs low for startups and SMBs yet scales cleanly for larger enterprises. In markets like automotive HMI, industrial inspection, and mobile UX automation, this balance of performance, footprint, and licensing could accelerate the shift from cloud-centric to on-device vision-language agents.
Source: MarkTechPost. AI-assisted editorial synthesis — TechnoExpress.