Liquid AI has released LFM2.5-VL-3B, its most capable vision-language model to date, built for on-device and edge deployment. The 3.1-billion-parameter model uses a non-reasoning architecture that answers directly rather than generating chains of thought, keeping latency low for real-time applications. The company says it delivers vision performance competitive with models twice its size while running faster than smaller rivals on CPU and GPU.
The release, announced Wednesday, builds on LFM2-VL-3B from October with four improvements. Screen and UI understanding is the flagship: the model averages 80.7 on ScreenSpot-v2, a benchmark of interface navigation, ahead of Gemma-4-E4B’s 51.2 and Qwen 3.5 4B’s 78.5, and close behind the larger InternVL-3.5-4B. Function calling, new to Liquid’s vision line, more than doubled on the ToolSandbox benchmark from 26.4 to 59.5. Grounding, the ability to locate objects described in natural language and return bounding boxes, jumped from 57.1 to 87.9 on RefCOCO precision at rank one. Multi-image reasoning also improved substantially, with BLINK rising from 50.2 to 61.5 and MUIRBench from 34.9 to 58.3.
Across 28 benchmarks spanning multilingual comprehension, instruction following, visual math and science, document understanding, detection, multi-image, and screen understanding, Liquid reports an average of 69.4, level with InternVL-3.5-4B and within 0.7 points of Qwen3.5-4B, both 4.7-billion-parameter models. The model leads its size class on RealWorldQA, POPE, and RefCOCO, covering physical-world questions, hallucination resistance, and object grounding, plus documents, charts, and UI elements.
Efficiency is the pitch that distinguishes it. The checkpoint runs in under 3.3GB of memory, decodes 228 tokens per second on an Apple M5 Max, and 116 tokens per second on an AMD Ryzen AI Max+ 395, with a phone-level result of about 20 tokens per second on a Galaxy S26 Ultra reported by third-party coverage. It ships in native, GGUF, ONNX, and MLX formats with day-one support for llama.cpp, MLX, and vLLM, and a WebGPU demo runs entirely in the browser, including webcam capture.
Training details explain part of the gain. The model uses the same architecture as Liquid’s smaller vision variants, pairing the LFM2.5-2.6B text backbone with a SigLIP2 400M vision encoder. Pretraining ran on roughly 34 trillion tokens with a vocabulary doubled to 128K to support non-Latin scripts, and vision pretraining was scaled fourfold with curated and synthetic data covering captions, OCR, grounding, and instruction following. Post-training combined supervised fine-tuning, including distillation from a larger teacher and the company’s Antidoom technique, with multi-reward reinforcement learning.
The model is available under the LFM Open License v1.0 on Hugging Face, with commercial use free below $10 million in annual revenue, and can be tried on Liquid’s playground. It is the third major release in the LFM2.5 family this month, following the 2.6B text model and the encoder line aimed at long-context CPU inference, as the MIT-born startup pushes a broader thesis: foundation models small and efficient enough to live on the devices people actually carry.
Sources: LFM2.5-VL-3B: A Better and Faster Vision-Language Model for the Edge (Liquid AI, Aug 12, 2026); LFM2.5-VL-3B for Better and Faster Vision Capabilities for the Edge (Hugging Face, Aug 12, 2026); Liquid AI Releases LFM2.5-VL-3B: A 3B Vision-Language Model That Reads Screens, Grounds Objects, and Calls Tools On-Device (MarkTechPost, Aug 13, 2026)