UI-Mate-9B-GGUF

UI-Mate-9B-GGUF

UI-Mate-9B is an open-weight foundation GUI agent from Tencent HY Frontier, built on Qwen3.5-9B, designed for long-horizon computer use across applications and operating systems by observing live screenshots and producing structured keyboard and mouse actions rather than relying on coordinate replay. Given a natural-language task instruction, screenshots, and interaction history, it outputs reasoning, a concise action description, and structured computer-use tool calls spanning mouse, keyboard, scrolling, waiting, user-interaction, and task-completion actions compatible with pyautogui, trained via supervised fine-tuning followed by online reinforcement learning in executable GUI environments. It's the mid-sized checkpoint in the UI-Mate family (alongside the larger UI-Mate-27B for general use and UI-Mate-democua-27B for demonstration-guided computer use), achieving strong results for its scale — 66.2 average score on OSWorld-Verified, 61.7 on WindowsAgentArena, and 34.00 strict success / 66.55 progress on OSWorkerBench — and is served via an OpenAI-compatible vLLM endpoint (retaining five screenshots of context by default) alongside a reference Python agent harness. Intended for research and development of screenshot-based GUI agents in controlled desktop environments, it requires an external runtime to execute predicted actions, is sensitive to application versions, screen layouts, and UI state changes, and Tencent recommends isolated/disposable environments, human confirmation before sensitive operations, and never treating model-reported success as proof of an achieved outcome; it's released under the Apache License 2.0.

[!NOTE] HF Papers: https://huggingface.co/papers/2608.15930

Model Files

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp