Liquid AI's LFM2.5-VL-3B: a screen-reading agent that runs in 3GB
The 3.1B open-weight VLM grounds objects, calls tools and OCRs docs — 228 tok/s on an M5 Max, ~3GB RAM. Plus: MAI-Code-1.1-Flash hits Copilot at 1/4 the cost.
Copy markdown
A 3B open-weight model that reads any screen
LFM2.5-VL-3B is a 3.1B vision-language model tuned for UI understanding across mobile, web and desktop: 80.7 on ScreenSpot-v2 (vs 51.2 for Gemma-4-E4B), object grounding at 87.9 on RefCOCO (up from 57.1), plus 91.1 on DocVQA and 84.2 on OCRBench. That's frontier-grade screen reading in weights you host yourself.
228 tok/s on a laptop, 20 on a phone, ~3GB RAM
It decodes 228 tok/s on an Apple M5 Max, 116 on an AMD Ryzen AI Max+ and 20 tok/s on a Galaxy S26 Ultra, fitting in about 3GB. Weights ship in four formats — native, GGUF, ONNX and MLX — with llama.cpp, MLX, vLLM and SGLang support, so it drops onto your hardware today.
Tool-calling doubled — build an on-device UI agent this weekend
Function-calling more than doubled (ToolSandbox 59.5, up from 26.4; BFCL v4 32.5, up from 20.5), so the model can see a screen and act on it with no cloud round-trip. Wire it into a local loop for an agent that navigates an app, fills forms or reads receipts — nothing billed, nothing leaving the device.
Read the $10M license line before you ship
It ships under the LFM Open License v1.0: free commercial use only if your company is under $10M in annual revenue; above that you need a negotiated license. Research, education and non-profit use face no revenue cap.
Elsewhere: MAI-Code-1.1-Flash lands in Copilot at a quarter the cost
Microsoft shipped MAI-Code-1.1-Flash to GitHub Copilot — a quarter the cost of June's 1.0, streaming 25% faster with 25% fewer tokens, and +22% on Terminal-Bench 2.1 for CLI work. It's off by default, so pick it in the model switcher; the older MAI-Code-1-Flash is being deprecated.