UI-Mate

Tencent Hy Frontier Team's open-weight foundation GUI agents ("Advancing Open-Weight Foundation GUI Agents with In-Context Demonstrations"): agent checkpoints post-trained on Qwen3.6-27B and Qwen3.5-9B with an environment-grounded training stack, plus a demonstration-guided variant (DemoCUA) that learns a user's workflow from in-context demonstrations. UI-Mate-27B reports 77.0 on OSWorld-Verified and 66.2 on WindowsAgentArena, which the paper calls the strongest open-weight result there, and 41.0 strict success (76.9 progress) on OSWorkerBench, a new office-centric benchmark released with the model: 100 long-horizon tasks across 17 business tracks and 41 applications with executable evaluators. Apache 2.0; the models are meant to run inside the repository's prompt, parser, and interaction harness rather than as chat models. Filed late: weights August 14, paper August 16, benchmark data September 1.