Muse Glimmer vs. Qwen 3.6 27B: Meta's New Local Model Doesn't Win Everything
Meta's Muse Glimmer wins on SWE-Bench Pro and AIME, but Qwen 3.6 27B wins on the practical coding benchmarks and uses less VRAM. Here's which one to actually run.
Contents6
Muse Glimmer vs. Qwen 3.6 27B: Meta's New Local Model Doesn't Win Everything
Meta put out Muse Glimmer on August 10, 2026, and the headline was the usual one: a 30-billion-parameter agentic model, Apache 2.0 licensed, small enough to run on a single consumer GPU. What the headline skips is that the model it's most directly compared against, Alibaba's Qwen 3.6 27B, still beats it on the benchmarks that matter most for day-to-day coding work. This is a split decision, not a sweep, and which model wins depends entirely on what you're using it for.
What Muse Glimmer is
Glimmer is Meta Superintelligence Labs' bet on always-on local agents rather than another chat model. The weights ship at roughly 4-bit precision, which shrinks the model itself to under 20 GB, leaving enough room in a 24 GB or 32 GB card for working memory, an image-understanding encoder, and a speculative decoding drafter to run at the same time. It's built for long-horizon reasoning, tool calling, failure recovery, and multimodal input, and it's available now on Hugging Face under Apache 2.0, with llama.cpp, MLX, and ExecuTorch integrations landing shortly after launch.
The benchmark split nobody's headline mentioned
Independent testing (not just Meta's own numbers) shows the two models trade wins depending on the benchmark:
| Benchmark | Winner | What it measures |
|---|---|---|
| SWE-Bench Pro | Muse Glimmer | Harder, less-contaminated software engineering tasks |
| AIME | Muse Glimmer | Competition math reasoning |
| SWE-Bench Verified | Qwen 3.6 27B | Practical, real-world coding fixes |
| TerminalBench | Qwen 3.6 27B | Command-line and terminal-driven agent tasks |
That split matters because SWE-Bench Verified and TerminalBench are the two benchmarks closest to what a coding agent does day to day: read an existing codebase, make a targeted fix, run it in a terminal, confirm it works. Glimmer's wins are concentrated in harder, more abstract reasoning and math, which is real but less relevant if you're running an agent against your own repo.
Hardware: where Qwen 3.6 27B pulls ahead
Qwen 3.6 27B is also the more efficient model to run. At Q4 quantization it fits in roughly 16.8 GB of VRAM, while Muse Glimmer needs closer to 18-20 GB once you count its perception encoder and speculative decoder. On a 24 GB consumer card (an RTX 4090 or similar), that difference is the gap between comfortable headroom and running right at the edge. If you're on a 16 GB card, Qwen 3.6 27B is the one that fits at all.
So which one should you run
Pick Muse Glimmer if you want a single local model for mixed agentic work, tool use, and reasoning-heavy tasks, and you have 24 GB or more of VRAM to spare. Pick Qwen 3.6 27B if your main use case is a coding agent that needs to pass practical, repo-level benchmarks, or your GPU tops out around 16 to 20 GB. Pick neither, and use a hosted model instead, if you don't already own the hardware. Renting equivalent GPU time usually costs more than a frontier API would for anything short of constant daily use.
Both models slot into local coding setups that already support model choice. [Kilo](https://questloops.com/tools/kilo), for instance, routes requests across 500+ models including local and BYOK options, so swapping between Glimmer and Qwen for different tasks doesn't mean switching tools. It's a similar story to what we saw with [Thinking Machines' Inkling-Small](https://questloops.com/blog/thinking-machines-inkling-small-explained-a-quarter-the-size-beating-its-own-bigger-sibling): smaller, locally-runnable models keep landing, and the right pick is task-specific, not universal.
The takeaway
Meta shipping an open-weights, agentic-tuned 30B model under Apache 2.0 is genuinely good news for anyone who wants to run agents locally without a subscription. Just don't take "Meta's new model" as shorthand for "the new best model." On the benchmarks closest to real coding work, a Qwen model roughly a tenth smaller still wins, and it asks for less VRAM while doing it.
