Muse Glimmer 30B vs Qwen3.6 27B: two ~30B open models compared

On August 10, 2026, Meta released Muse Glimmer 30B under Apache 2.0, pitched as a local-agent model that runs on a single consumer GPU. The model Meta chose to measure it against is Qwen3.6 27B, released by Alibaba in April. Both are on FastMetal, so you can run them side by side on one key.

The short answer: Qwen3.6 27B when coding accuracy comes first; Muse Glimmer 30B for running agent loops cheaply. Glimmer's output costs about half of Qwen's, and its cached-input rate widens the gap in loops that resend the same context.

Specs and price

Prices are what a USD account pays today; current figures are on the pricing page. Input costs about the same on both. Output is where they split: Qwen's is a little over twice Glimmer's.

Benchmarks, and how to read them

Meta reports a score of 65.1 on the Beam128K agent benchmark, ahead of Gemma 4 31B and Qwen3.6 27B. Read the same release closely, though, and Qwen comes out ahead on several of the coding and computer-use evaluations in Meta's own tables. On the third-party BenchLM aggregate, Muse Glimmer scored 52.31 at launch, 110th of 218 models: mid-pack, not top of its class.

Qwen3.6 27B launched with the claim of flagship-level coding from a 27B dense model, with repository-scale code understanding and front-end work as its strengths (both the vendor's claims). It also has four more months of production use behind it, and that track record counts.

Weak spots

• Muse Glimmer 30B: behind Qwen on some coding evaluations, even in Meta's comparison. Reasoning is always on and can't be turned off, so even short answers include thinking tokens in the output count; when we measured it at launch, 44 of 51 output tokens were reasoning. Its context window is half of Qwen's. And it is a new model, with little production experience reported yet.
• Qwen3.6 27B: output costs about twice Glimmer's, and there is no cached-input discount on this route, so loops that resend the same context get no relief.

What one task costs

Take one sizable task: 200,000 input tokens and 20,000 output tokens, the scale of handing a model a chunk of code or documents and having it work.

That is about 1.3 times. The difference grows in an agent loop. If three quarters of those 200,000 input tokens are cache hits on Glimmer, its cost drops to about $0.057 per task, while Qwen's stays at $0.135.

Which one to use

• Muse Glimmer 30B: multi-turn agent loops, cost-sensitive routine tasks, and testing with a possible move to local hardware later, since the weights are open.
• Qwen3.6 27B: code edits and generation where accuracy matters, contexts beyond 131K tokens, PDF input, or a model with a longer track record.
• A practical setup: let Glimmer run the loop, and hand the hard code fixes to Qwen3.6 27B, or to a larger model.

Try both on one key

Only your own prompts will tell you which fits. Same key, same balance; only model changes:

curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "muse-glimmer-30b", "messages": [{"role": "user", "content": "Hello"}]}'

curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "qwen3.6-27b", "messages": [{"role": "user", "content": "Hello"}]}'

FAQ

Which one is better? On coding evaluations, Qwen3.6 27B. On agent-oriented design and price, Muse Glimmer 30B. The published benchmarks don't settle it, so compare them on your actual tasks.

Can Muse Glimmer 30B run locally? The weights are open under Apache 2.0, and Meta says a 4-bit quantization fits on a 24 GB GPU. FastMetal serves it hosted, so you can try the same model without any setup.

How do they compare across languages? Muse Glimmer claims support for more than 100 languages, Qwen3.6 for 201. We found no published benchmarks for either in specific languages, so test on your own prompts if you work outside English.

Summary

Muse Glimmer 30B isn't a model you pick on benchmark scores. You pick it for the cost of an agent loop and a design built for one. Qwen3.6 27B is still stronger on first-try coding accuracy and has the longer track record. Glimmer for volume and loops, Qwen for the hard parts is the setup that wastes the least today.

Details are on the Muse Glimmer 30B and Qwen3.6 27B pages, and the full list is in the model catalog.