Repository · read 2026-10-17
meta-oss-cookbook
The cookbook is Meta's first-party answer to the question every open-weight release raises: which checkpoint, which runtime, and how much memory. The Muse Spark folder carries a quickstart, an inference-server guide, hosted-provider notes, prompting examples, post-training recipes and end-to-end walkthroughs including Mac Studio clusters.
The repository → 107 stars · Python
| Field | Value |
|---|---|
| Repository | meta-models/meta-oss-cookbook 107 stars · read 2026-10-17 |
| Language | Python Last push 2026-10-15 |
| Licence | Apache-2.0 (repository); the weights carry the Muse Model License Repository licence; the model weights carry their own. |
| Stack | Python recipes; SGLang, vLLM, MLX, llama.cpp, TensorRT-LLM What it runs on. |
| Best for | Anyone planning a self-hosted deployment who wants the vendor's own numbers before trusting a tutorial. A reader, not a benchmark. |
How it works
Each folder is a set of commands and configuration rather than a library. The model README fixes the two checkpoint formats — BF16 at roughly 1.65 TB of weights and NVFP4 at roughly 537 GB — and the inference-server folder points at SGLang and vLLM with the settings Meta recommends. The recipes folder holds the Mac Studio walkthrough.
What we checked
- The repository README table naming Muse Spark Open and Muse Glimmer
- muse-spark/README.md: parameter count, context window, checkpoint sizes, hardware, licence links
- The folder listing under muse-spark/
- The quickstart's stated BF16 and NVFP4 footprints
What we did not check
We did not download the weights or run any command in the repository; the GPU and Mac Studio hardware it assumes is not available here.
Our note
Meta's own recipes for running its open-weight models, and the only place with a checked set of Muse Spark Open commands: two checkpoint formats, the hardware each one needs, and the served API path. Start here before any blog post.
github.com/meta-models/meta-oss-cookbook Verified 2026-10-17 Stars 107 at read time