Muse Glimmer
2 MLX aliases · Meta's Muse Glimmer 30B · dense reasoner, 131K context.
Meta's Muse Glimmer — a 29.6B dense reasoning model with a sliding-window attention stack and recipient-routed channels: the model thinks on a private channel and answers on another, and rapid-mlx demultiplexes that wire natively into reasoning_content and content. Tool calls use the ATEM function-calls envelope, parsed by the dedicated muse parser pair. Since 0.12.9, rapid-mlx serves the text backbone through its own vendored implementation — no external runtime dependency. The checkpoint ships a vision tower, but image input stays off until the multimodal path lands upstream; text, reasoning and tool calls are fully supported today.
- family
- Muse Glimmer
- aliases
- 2
- lines
- 1
- install
- rapid-mlx serve <alias>
- OpenAI base URL
- http://localhost:8000/v1
Muse Glimmer 30B · 2 aliases
Dense 29.6B with [sliding ×3, global] attention, 131K context and channel-routed reasoning. Uses the muse tool-call parser (ATEM envelope) and the muse reasoning parser — <think>-free chain-of-thought streams out of reasoning_content with no extra config.
parser: muse
| alias | hf repo | tool parser | reasoning | flags |
|---|---|---|---|---|
| muse-glimmer-30b-4bit | mlx-community/Muse-Glimmer-30B-4bit | muse | muse | — |
| muse-glimmer-30b-bf16 | mlx-community/Muse-Glimmer-30B-bf16 | muse | muse | — |
Notes & caveats
- muse-glimmer-30b-4bit (19.4 GB) is the pick — fits a 24 GB+ Mac and is mirrored on our edge CDN.
- bf16 needs ~96 GB unified memory — Mac Studio territory.
- Vision weights ship in the checkpoint but image input is not served yet — text, reasoning and tool calls only.