UI-Mate-27B GGUF
Community GGUF quantization of
Tencent/UI-Mate-27B — an open-weight foundation GUI agent based on Qwen3.6-27B. It observes live screenshots, reasons over the visible state, and produces structured keyboard/mouse actions for native desktop interaction (pyautogui-compatible).
Converted with llama.cpp b10437 (convert_hf_to_gguf.py --no-nextn), quantized with llama-quantize (CUDA). The vision projector (mmproj) is included.
Files
| File | Size | Notes |
|---|
UI-Mate-27B-Q4_K_M.gguf | 15.4 GB | Recommended for 24 GB GPUs (RTX 3090/4090) |
UI-Mate-27B-Q5_K_M.gguf | 17.9 GB | Middle ground; 24 GB GPUs |
UI-Mate-27B-Q6_K.gguf | 20.6 GB | Highest fidelity; needs 24 GB+ with modest context |
mmproj-UI-Mate-27B-F16.gguf | 0.86 GB | Vision projector — required for image input |
Usage (llama.cpp)
Requires llama.cpp b10437+ (qwen3_5 / Gated DeltaNet support).
1llama-server \
2 -m UI-Mate-27B-Q4_K_M.gguf \
3 --mmproj mmproj-UI-Mate-27B-F16.gguf \
4 --ctx-size 65536 \
5 --n-gpu-layers 999 \
6 --image-min-tokens 1024 \
7 --n-predict 8192
Works with any OpenAI-compatible client (llama.cpp server, llama-swap, LM Studio, etc.). For full GUI-agent behavior (structured actions, coordinate rescaling, demonstration-guided mode), use Tencent's
UI-Mate harness against the endpoint.
Benchmarks (RTX 3090, CUDA, ngl 999)
| Metric | Q4_K_M |
|---|
| pp512 | 1341 t/s |
| tg128 | 42.3 t/s |
| VRAM (Q4_K_M + mmproj, 64K ctx) | ~19.7 GB |
Perplexity (wikitext-2, 10K tokens, ctx 512)
| Quant | PPL |
|---|
| Q4_K_M | 6.5171 ± 0.216 |
| Q5_K_M | 6.4924 ± 0.216 |
| Q6_K | 6.4663 ± 0.214 |
Monotonic improvement with precision; Q6 is ~0.05 PPL better than Q4 — quantization loss is minimal.
Community Validation (llama.cpp / llama-swap)
Tested with Tencent's official
UI-Mate harness against a llama.cpp b10437 server (llama-swap) on an RTX 3090 (24 GB).
- Compatibility: llama.cpp accepts the harness's
chat_template_kwargs.enable_thinking; the model emits both content (XML actions) and reasoning_content.
- Single-step (5/5): all bundled examples produced valid
<action> + <tool_call> output (e.g., Chrome bookmark → Ctrl+D hotkey; LibreOffice → File menu; Thunderbird → app icon).
- Replay (5 steps): 2/5 exact match to recorded actions, 4/5 within 1–2 px.
- Multi-turn: screenshot history + collapsing work correctly.
Note: Q4_K_M validated on 24 GB. Q6_K is provided for 24 GB+ GPUs (not GPU-validated on 24 GB).
Model Details (from the official card)
- Parameters: 27B
- Base model: Qwen3.6-27B
- Input: task instruction, screenshots, interaction history, and optional demonstration context
- Output: reasoning, a concise action description, and structured computer-use tool calls
- Action space: mouse, keyboard, scrolling, waiting, user interaction, and task completion
- Training: supervised fine-tuning followed by online reinforcement learning in executable GUI environments
- License: Apache-2.0
UI-Mate supports two complementary modes:
- General computer use: execute tasks from natural-language instructions and live screenshots.
- Demonstration-guided computer use: adapt a reusable workflow extracted from one successful demonstration to a new task.
Evaluation (from the official card)
Instruction-only execution
| Benchmark | UI-Mate-27B |
|---|
| OSWorld-Verified · average score | 77.0 |
| WindowsAgentArena · average score | 66.2 |
| OSWorkerBench · strict success | 41.00 |
| OSWorkerBench · progress | 76.86 |
Demonstration-guided execution (OSWorkerBench-Subset, 33 tasks)
| Metric | Instruction only | + one demonstration |
|---|
| strict success | 17.17 | 35.35 (+18.18 pp) |
| progress | 67.85 | 81.14 (+13.29 pp) |
Intended Use and Limitations
UI-Mate-27B is intended for research and development of screenshot-based GUI agents in controlled desktop environments. Its behavior can be affected by application versions, screen layouts, display scaling, latency, and unexpected UI state. Benchmark performance does not guarantee reliable execution in arbitrary environments, and the model requires an external runtime to execute its predicted actions.
Safety
Computer-use agents can make mistakes, encounter prompt injection, or trigger consequential actions.
- Prefer isolated or disposable environments.
- Avoid unattended, high-stakes, or destructive workflows.
- Require human confirmation before sensitive operations.
- Monitor the interaction trajectory and verify the resulting application state.
- Do not treat a model-reported success as proof that the intended outcome was achieved.
Notes
- Architecture:
qwen35 (Qwen3.5-family Gated DeltaNet hybrid), 64 layers, hidden 5120, native 262K context. MTP head excluded from this conversion (--no-nextn).
- Reasoning model: responses include
reasoning_content before the final answer — set a generous max_tokens (e.g. 8192) for long-horizon tasks.
- Grounding:
--image-min-tokens 1024 is recommended for GUI grounding accuracy (llama.cpp warning for Qwen-VL models).
- Original model: Apache-2.0. Third-party components retain their licenses.
Credits
- Model: Tencent HY Frontier (UI-Mate) — repo · project page
- Base: Qwen/Qwen3.6-27B
- Quantization: llama.cpp b10437 (official toolchain)
Citation
1@article{uimate2026,
2 title = {UI-Mate: Advancing Open-Weight Foundation GUI Agents with In-Context Demonstrations},
3 author = {{Tencent HY Frontier Multimodal Agent Team}},
4 journal = {arXiv preprint},
5 year = {2026}
6}
