UI-Mate 27B vs Qwen 3.6 27B

Analysis

UI-Mate 27B and Qwen 3.6 27B are, underneath, the same 27-billion-parameter checkpoint. UI-Mate 27B is Tencent's GUI-agent fine-tune built directly on Qwen 3.6 27B's weights, so its architecture, weight size, and KV-cache math are identical byte-for-byte, confirmed by UI-Mate's own published config.json. That makes this comparison unusual for this site: there's no VRAM tradeoff to analyze at all, so the only real question is what job each version of the weights was trained to do.

Both models share the exact 64-layer hybrid stack, 16 Gated Attention layers (4 KV heads x 256 head dim) interleaved with 48 Gated DeltaNet linear-attention layers, the same 262,144-token native context window, and the same Apache 2.0 license. At any quant or context length, their weights and KV-cache totals are identical: 16.4 GB at Q4_K_M, 19.0 GB total at 8k context, 37.7 GB at the full 262,144-token window. Capability is where the two diverge completely. Qwen 3.6 27B was trained and evaluated as a general chat/coding/reasoning model: Alibaba reports MMLU-Pro 86.2, GPQA Diamond 87.8, and SWE-bench Verified 77.2. UI-Mate 27B went through supervised fine-tuning and agentic reinforcement learning on desktop-automation trajectories instead, and Tencent evaluates it on a completely different set of benchmarks: OSWorld-Verified (77.0), WindowsAgentArena (66.2), and Tencent's own OSWorkerBench suite, where it scores 41.0% strict success and 76.9% progress against 23.3% and 52.4% for the untouched Qwen 3.6 27B base on the same 100 tasks (Tencent's own ablation, arXiv:2608.15930). Neither vendor reports the other model's benchmark suite, so the comparison table above shows real numbers for Qwen 3.6 27B and blanks for UI-Mate; that isn't missing data, it reflects that UI-Mate was never evaluated as a general-purpose model at all. UI-Mate also loses something in the trade: at launch it ships only as BF16 safetensors, no GGUF, no Ollama tag, so llama.cpp and Ollama users can run Qwen 3.6 27B today but have nothing to load for UI-Mate until a community quant appears, while Qwen 3.6 27B's own GGUFs have been through months of community requantization by comparison.

Bottom line: If you want a general-purpose local coding or chat model, run Qwen 3.6 27B directly: it's the more capable and more immediately usable of the two, with mature GGUF support and identical hardware requirements. Run UI-Mate 27B only if your actual task is driving a desktop GUI from screenshots, since that's the one thing its fine-tuning specifically bought it over the base checkpoint, and only once a community GGUF exists or you have access to an 80GB+ card for the native BF16 weights through vLLM. There's no VRAM reason to prefer one over the other: whichever you pick, the hardware bill is the same.