Meta's Muse Glimmer model - smoke testing

TLDR: Muse Glimmer is slower than my local coding agent baseline

Company Meta’s announcement yesterday: Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device.

I was curious to swap out the Qwen 3.6 model with Muse Glimmer in my local-coding-agent framework.

Test environment

• MacBook Pro (M5 Pro, 48GB)
• Ollama version 0.32.8
• Ollama model tags:
  • muse-glimmer:30b-mlx
  • baseline: qwen3.6:35b-mlx with modified sampling parameters

First I ran Sebastian Raschka’s benchmarking script and saw Muse Glimmer run more than 3x slower at around 20 tokens per second.

I then asked it to run a repo code review for a macOS/iOS app I am developing. After 10 minutes of thinking the Qwen Code API timed out. I’ll try later to see if Ollama implementation of Muse Glimmer runs faster.

Conclusion: I’ll stick with the Qwen model for local coding work.