Qwen 3.8 27B vs Muse Glimmer 30B

Analysis

Qwen 3.8 27B and Muse Glimmer 30B are both dense, sub-30B agentic models that shipped four days apart in August 2026, Muse Glimmer on the 10th, Qwen on the 14th. The timing matters for more than trivia: Muse Glimmer's own launch benchmarks were measured against Qwen 3.6 27B and Gemma 4 31B, the two same-week peers Meta had on hand at the time. Qwen 3.8 27B, which replaced Qwen 3.6 27B the same week Muse Glimmer shipped, was never part of that comparison, so this pairing is the first place the two models actually sit side by side.

Qwen 3.8 27B kept its predecessor's hybrid stack unchanged: 64 layers, 16 of them Gated Attention (4 KV heads, 256 head dim) that keep a real KV cache, the other 48 running Gated DeltaNet, linear attention whose memory is a fixed-size recurrent state that doesn't grow with context. Muse Glimmer 30B reaches for the same goal a different way: all 52 of its layers are ordinary attention, but 39 of them, a repeating pattern of three sliding-window layers per one full-attention layer, cap their cache at a 2,048-token window, paired with one of the narrowest KV widths tracked on this site (2 KV heads, 128 head dim, a 16:1 GQA ratio). Qwen's native context window, 262,144 tokens, is double Muse Glimmer's 131,072, so the two only overlap up to Muse Glimmer's own ceiling. At that 131,072-token limit, Qwen'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 Qwen 3.8 27B against 21.0 GB for Muse Glimmer 30B, close to 7 GB apart on the same card. On the benchmark fields both models report, each vendor's own numbers, not a controlled shared test, Qwen leads clearly: GPQA Diamond 89.2 against 83.5, SWE-bench Pro 61.7 against 51.2, and Terminal-Bench 2.1 73.0 against 51.7, more than double Muse Glimmer's score on that last one. Modality splits the other way in scope, not favor: Qwen 3.8 27B natively understands video as well as images, where Muse Glimmer's vision input, a separate ViT-G/14 encoder shipped as its own mmproj file, stops at still images. Both ship under an unrestricted Apache 2.0 license. Neither had a finished day-one tooling story: Qwen 3.8 27B shipped with no Ollama tag at all, and Muse Glimmer's only launch tag, muse-glimmer:30b-mlx, runs exclusively through Apple's MLX engine, with NVIDIA and AMD support still rolling out as of release day. llama.cpp or vLLM is the dependable route for either model on non-Apple hardware.

Bottom line: Qwen 3.8 27B is the stronger pick on raw benchmark scores and context length: it leads Muse Glimmer on every metric both report and reaches twice the context window. Muse Glimmer 30B's case is narrower: pick it for the smaller VRAM footprint, about 7 GB less at the 131,072-token ceiling both models can reach, or if you're specifically on Apple Silicon, where Muse Glimmer's day-one MLX tag already works and Qwen's Ollama support doesn't exist yet.