Published: 18 August 2026
Tencent released UI-Mate-27B on 17 August 2026, an open-weight agent that operates a desktop from screenshots and scores 77.0 on OSWorld-Verified by Tencent’s own testing, above Kimi-2.6 at 73.1. A smaller UI-Mate-9B and a document retrieval model, EVIE-Preview-4.5B, shipped alongside it. All three are Apache 2.0 with no hosted API.
What can UI-Mate-27B do?
UI-Mate-27B reads live screenshots of a desktop and returns keyboard and mouse actions in a pyautogui compatible format, working across Ubuntu, Windows and macOS. It is post trained from Alibaba’s Qwen3.6-27B; UI-Mate-9B uses Qwen3.5-9B. Neither is a chat model, and both need an external runtime to actually execute the actions they predict.
The distinctive feature is in context demonstrations. Supply one recorded example of a task alongside the instruction and Tencent reports strict success on a 33 task office benchmark rising from 17.2 to 35.4 percent. EVIE-Preview-4.5B is unrelated work: a retrieval model that encodes page images into 128 dimension embeddings for document heavy RAG.
UI-Mate-27B benchmarks and how far to trust them
Tencent reports UI-Mate-27B at:
- 77.0 on OSWorld-Verified,
- 66.2 on WindowsAgentArena,
- 41.0 strict success on OSWorkerBench.
- UI-Mate-9B reaches 66.2, 61.7 and 34.0.
- EVIE-Preview-4.5B averages 64.40 nDCG@10 on the public ViDoRe V3 split, just ahead of nemotron-colembed-vl-8b-v2 at 63.54.
All of these are Tencent’s own runs through its own harness, and no third party has reproduced them. Two figures deserve more caution than the rest: OSWorkerBench is a new benchmark Tencent created and has not yet released, and the GitHub README states that final public benchmark numbers are still being verified. The OSWorld-Verified and WindowsAgentArena scores are at least checkable in principle.
How does UI-Mate-27B compare to Kimi-2.6 and Qwen3.6-27B?
Against Kimi-2.6, a 1 trillion parameter model with 32B active, Tencent puts UI-Mate-27B 3.9 points ahead on OSWorld-Verified and 2.9 ahead on WindowsAgentArena. If that holds up, a 27B model is matching a frontier system on computer use at a fraction of the serving cost.
The more telling comparison is with its own base. Qwen3.6-27B scores 52.5 on OSWorld-Verified and 23.3 strict success on OSWorkerBench, so almost all of the capability comes from Tencent’s post training rather than the underlying model. Both comparison sets are published by Tencent.
UI-Mate-27B availability, licence and hardware
All three models are on Hugging Face, ungated, under Apache 2.0. There is no hosted API and no first party pricing, so running them is the only option. UI-Mate-27B needs roughly 55 GB of memory for the BF16 weights, putting it on a single 80 GB accelerator; UI-Mate-9B fits in about 19 GB.
Tencent also ships a free macOS client for UI-Mate, Apple Silicon only. Note that neither UI-Mate model executes anything by itself: you supply the runtime that turns predicted actions into real clicks, which is where sandboxing and approval policies have to live.
What this means
If you are building desktop automation, UI-Mate-27B is worth a pilot this quarter. It is the first open-weight checkpoint with a credible claim to frontier-level computer use, and Apache 2.0 plus local weights means you can point it at internal systems you would never expose to a hosted API.
Do not put it near a customer or an unattended workflow yet. Tencent’s own card warns against high stakes and destructive tasks, and its strongest numbers rest on a benchmark nobody outside Tencent can run. EVIE-Preview-4.5B is the lower risk of the two: a 128 dimension index cuts document RAG storage sharply, and you can test retrieval quality against your own corpus in an afternoon.
Model cards, code and the technical report are published at the Tencent Github and on the Tencent Hugging Face page.
