Liquid AI’s LFM2.5-VL-3B is a 3.1‑billion‑parameter vision‑language model built for on‑device use. It solves the common problem of needing strong screen and UI understanding without sacrificing speed or memory limits. Developers often struggle with models that are too large for mobile or edge hardware, or that require separate reasoning steps that add latency. LFM2.5-VL-3B answers directly, keeping latency low while still grounding objects to coordinates, parsing documents and charts, and calling tools from text or image input.
The model fits in roughly 3 GB of memory and decodes about 228 tokens per second on an Apple M5 Max. It ships in four ready‑to‑run formats—native, GGUF, ONNX, and MLX‑8bit—so you can deploy with llama.cpp, MLX, vLLM, SGLang, or ONNX runtimes without extra conversion work. Its open‑source‑based LFM Open License v1.0 allows free commercial use for companies under $10 M annual revenue, making it practical for indie developers, startups, and SMBs. Enterprises above that threshold simply negotiate a commercial license.
Performance wise, LFM2.5‑VL‑3B scores 69.4 average across 28 vision benchmarks, matching InternVL‑3.5‑4B and trailing Qwen3.5‑4B by only 0.7 points despite being far smaller. It improves screen understanding (80.7 on SpotSpot‑v2), grounding (RefCOCO precision@1 up to 87.9), multi‑image input, and function calling, enabling use cases such as on‑device screen agents, GUI test automation, invoice OCR, near‑real‑time object detection in vehicles, offline translation of menus and signs, and multi‑image comparison.
To get started, download the desired format from Hugging Face, load it with your preferred runtime, and feed screen captures or images directly. The model’s non‑reasoning design means you receive answers instantly, keeping your application responsive.
#AI #MachineLearning #OnDeviceAI #VisionLanguage #EdgeComputing #Productivity