ReviewedOfficial evidence
UI-Mate-27B
Official Deployment Path and Safety BoundariesA source-based profile of Tencent’s 27B open-weight GUI agent, including its screenshot-driven action loop, official vLLM deployment path, runtime requirements, and computer-use safety boundaries.
Evidence note
This page organizes first-party documentation and editorial analysis. ModelRun Lab has not independently reproduced the published performance or hardware claims.
Model overview
UI-Mate-27B is an open-weight GUI agent published by Tencent HY Frontier. It is designed for long-horizon computer-use tasks across applications and operating systems. The model consumes task instructions, live screenshots, and interaction history, then returns reasoning, a concise action description, and structured computer-use tool calls.
UI-Mate is an agent checkpoint, not a standalone visual-chat application. It requires an external interaction harness to execute predicted mouse, keyboard, scrolling, waiting, and user-interaction actions. ModelRun Lab has not independently run the checkpoint or reproduced Tencent’s benchmark results.
Published model details
Tencent lists the following details:
- 27 billion parameters
- Qwen3.6-27B base model
- Screenshot, task instruction, and interaction-history inputs
- Structured computer-use tool-call output
- Supervised fine-tuning followed by online reinforcement learning
- PyAutoGUI-compatible actions
- OpenAI-compatible serving and client interface
- Apache 2.0 repository license metadata
The provider positions UI-Mate-27B for research and development in controlled desktop environments.
Official deployment path
The official model card documents serving with vLLM. Its example uses tensor parallelism across two devices, GPU memory utilization of 0.85, multimodal encoder tensor-parallel data mode, and admission for up to six images per prompt.
The six-image setting supports the default agent behavior, which retains five screenshots while the newest screenshot enters context. This is an official example configuration, not a universal minimum hardware specification.
Tencent also provides an official repository with a reference agent, response parser, recorded-trajectory example, and an OpenAI-compatible client. The model card recommends using that prompt, parser, and interaction harness rather than treating the checkpoint as a generic chat model.
Hardware boundary
The reviewed model card provides a two-device vLLM example but does not identify GPU models or publish one universal minimum VRAM figure. Memory requirements vary with precision, runtime version, screenshot count and resolution, context history, tensor parallelism, and serving overhead.
This page therefore does not convert the example into a consumer-GPU guarantee or claim that ModelRun Lab has tested the deployment.
Provider-reported evaluation
Tencent reports results on OSWorld-Verified, WindowsAgentArena, and OSWorkerBench. Those scores are provider-reported and should be assessed with the project page and paper methodology. Benchmark performance does not prove reliable execution in a different application version, display layout, latency environment, or workflow.
Safety boundary
Computer-use agents can take consequential actions in external applications. Tencent’s model card warns about errors, prompt injection, and destructive or sensitive operations. Its published precautions include:
- Prefer isolated or disposable environments
- Avoid unattended, high-stakes, or destructive workflows
- Require human confirmation before sensitive actions
- Monitor the interaction trajectory
- Verify the resulting application state independently
A model-reported success is not evidence that the intended task completed correctly. Any deployment should enforce permissions and confirmation outside the model.
License boundary
The model card labels UI-Mate Apache 2.0 and notes that third-party components retain their own licenses. This page reports the repository metadata and does not provide legal advice. Review the current repository license, base-model terms, dependencies, and any applicable policies before use.
Evidence boundaries
- Architecture, official runtime commands, intended use, and safety guidance come from Tencent’s official model card and repository.
- Evaluation results are provider-reported.
- No universal minimum GPU or VRAM claim is made.
- ModelRun Lab has not independently tested the agent.