Qwen-UI-Agent: open-weights agents that drive your phone and PC
Alibaba's Tongyi lab open-sources a GUI-agent family from 2B to 235B, Apache 2.0 — plus Autolith's live-runtime agent and A2A joins the Linux Foundation.
Copy markdown
A computer-use agent you can actually download
Tongyi-MAI released Qwen-UI-Agent (the MAI-UI family) as open weights in four sizes — 2B, 8B, 32B and 235B-A22B, all Apache 2.0 on Hugging Face. It acts from raw screen pixels, so it can drive apps that expose no API at all.
Build this weekend: a phone-driving agent on one GPU
The 2B and 8B checkpoints run on a single consumer GPU via vLLM behind an OpenAI-compatible endpoint — enough to script an agent that fills forms, books, or navigates any Android app by tapping the UI. No device farm or vendor API required.
The leaderboard receipts — on the authors' evals
Qwen-UI-Agent tops ScreenSpot-Pro grounding at every size (67.9% at 32B, 65.7% at 8B), and the 235B takes #1 on AndroidWorld at 76.7% with a claimed new MobileWorld record. These are the team's own numbers, so benchmark on your own workflow before betting on it.
Elsewhere: Autolith gives a coding agent a live Lisp runtime
Autolith 0.35 (open source, Nix-installable) is a terminal coding agent that lives inside an SBCL image it can inspect, patch and crash-recover mid-task, and it processes corpora larger than the context window. Bring your own provider — Codex, Anthropic, Grok or Fireworks.
Elsewhere: A2A moves in with MCP under one foundation
Google donated its Agent2Agent protocol to the Linux Foundation's Agentic AI Foundation, which now houses A2A alongside Anthropic's MCP across 250+ members. That means fewer competing wire formats for your agents to speak, under neutral governance.