Ant's open-source UI-Venus-2: 9B model beats larger LLMs across phone, desktop and web GUI tasks
大模型之路 · wechat · 2026-09-17
Ant Group's InclusionAI released UI-Venus-2, an open-source (Apache-2.0) cross-platform GUI agent where one checkpoint handles mobile, desktop and web. The 9B version scores 90.8% on WebVoyager (vs Kimi-K2.6's 76.8%) and 65.8% on MobileWorld; the 27B version hits 93.4% on WebVoyager and 84.0% on AndroidWorld, setting several SOTAs including CAPTCHA benchmarks.
Key techniques:
- Three-stage training: large-scale trajectory mid-training, scenario-specific offline RL, then multi-teacher online distillation (MOPD) with action-aware supervision
- Two-level verification (trajectory-level + sample-level) with heterogeneous model voting to avoid mislabeling partial completions as success
- Ships with Venusframeworkmobile (40+ Chinese apps) and a browser plugin; safety evals on OSHARM/OS-BLIND
Caveats: benchmarks are vendor-reported and not independently reproduced yet.
More from coding & agent
- Tencent unveils Hyra 2.0, a research agent that builds reusable scientific tools — TencentHunyuan · 2026-11-02
- What broke when I ran a leaderless agent swarm with no orchestrator — 7 failure modes and fixes — LostHumanity67 · 2026-11-02
- Forced yes/no prompts push GPT and Claude to 95%+ confidence on Goldbach's unproved halt test — Nir777 · 2026-11-02
- Automation took over my daily work — so why am I still sitting in the office for 8 hours? — Am1rSori · 2026-11-02
- General Reasoning explains how to train agent swarms with RL and latency pressure — rosstaylor90 · 2026-11-02
- "Coding isn't the bottleneck anymore — usage windows are," says founder Kyle Gawley — kylegawley · 2026-11-02