UI-Mate 27B vs Muse Glimmer 30B

Analysis

UI-Mate 27B and Muse Glimmer 30B are both agent-focused local releases from August 2026, six days apart (Muse Glimmer on 10 August from Meta, UI-Mate on 16 August from Tencent), but built for different jobs: UI-Mate is a dedicated GUI-automation agent fine-tuned to read screenshots and drive a desktop, while Muse Glimmer is a general tool-calling and agentic-workflow model. Both also solve the "keep the KV cache small" problem, using completely different tricks: UI-Mate inherits Qwen 3.6 27B's linear-attention stack, Muse Glimmer uses Meta's own sliding-window design.

UI-Mate 27B carries Qwen 3.6 27B's hybrid stack over unchanged: 64 layers, 16 of them (4 KV heads x 256 head dim) real Gated Attention with a growing KV cache, the other 48 Gated DeltaNet layers running a fixed-size recurrent state instead. Muse Glimmer 30B takes a different approach: every one of its 52 layers is real attention, but 39 of them (a repeating 3-local/1-global pattern) cap their cache to a 2,048-token sliding window, combined with a narrow 32-query/2-KV-head ratio. At 131,072 tokens, the context length where Muse Glimmer's native window ends (UI-Mate's continues to 262,144, inherited from Qwen 3.6 27B), UI-Mate's KV cache is 8.6 GB against Muse Glimmer's 1.8 GB. Add each model's Q4_K_M weights and the full picture is 28.0 GB total for UI-Mate against 21.0 GB for Muse Glimmer, past a single 24 GB card for UI-Mate at that context length, comfortably inside it for Muse Glimmer. At a more typical 8,192-token context the gap nearly disappears: 19.0 GB for UI-Mate against 19.2 GB for Muse Glimmer, functionally identical. Capability isn't a fair fight in either direction, because the two were trained for different jobs and evaluated on different suites: Muse Glimmer reports general benchmarks (GPQA Diamond 83.5, SWE-bench Verified 76.0, SWE-bench Pro 51.2, Terminal-Bench 2.1 51.7) plus Meta's own agentic-eval table (MCP Atlas, DeepSearch QA, Gaia2), none of which Tencent evaluated UI-Mate on; UI-Mate reports OSWorld-Verified (77.0), WindowsAgentArena (66.2), and its own OSWorkerBench (41.0% strict success), none of which Meta evaluated Muse Glimmer on. The two do overlap on one suite, loosely: Meta's own comparison table measured Muse Glimmer at 65.9 on OSWorld-Verified, and Tencent's own figure for UI-Mate on the same suite is 77.0, but these are two different vendors' separate harnesses on two different checkpoints, not a shared measurement, so treat the gap as directional rather than exact. Licensing is identical (Apache 2.0, no usage restrictions either way), and both are text-plus-vision, neither handles video. Tooling favors neither cleanly at launch: Muse Glimmer's only day-one Ollama tag is Apple Silicon (MLX) only, and UI-Mate shipped with no GGUF or Ollama build of any kind, only BF16 safetensors for vLLM.

Bottom line: These two aren't really competing for the same job. If you need a general local agent for tool use, search, or long-running assistant workflows, Muse Glimmer 30B is the more practical pick today: real GGUF and Ollama support are already forming, its footprint stays smaller at long context, and it reports scores across a broad agentic suite. If your task is specifically driving a desktop GUI from screenshots, UI-Mate 27B is the only one of the two actually built and evaluated for it, but budget for the fact that there's no consumer-GPU path yet: an 80GB+ card through vLLM is what it takes until a community GGUF appears.