Builds · Muse Spark Open
Built with Muse Spark Open
The first 48 hours produced one official cookbook and three experiments, each aimed at a different machine: two Mac Studios' worth of MLX, two DGX Sparks at a 1-bit quant, and a 128 GB laptop streaming experts from its SSD. There is also an independent benchmark that asks whether the model can review code, which is the claim the launch leaned on.
Stars and last push read 2026-10-17 4 described, each with a page of its own Search found no other repository naming the model on 2026-10-17
- GitHub
GitHub
meta-oss-cookbook
Official recipes for serving and post-training Muse Spark Open: BF16 and NVFP4 checkpoints, SGLang and vLLM, hosted providers, prompting and fine-tuning.
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
GitHub
mlx-moe-ssd-streaming
An MLX layer that runs a 346 GB quantized Muse Spark Open checkpoint on a 128 GB Mac by caching 25% of experts in memory and streaming the rest from the internal SSD.
The most interesting experiment in the first 48 hours: keep 17 GB of dense weights in RAM, stream the routed experts from an SSD, and run an 816B model on a 128 GB MacBook at about 4.4 tokens per second.
- GitHub
GitHub
muse-spark-open-2x-dgx-spark
A reproducible two-node DGX Spark recipe for Muse Spark Open using llama.cpp RPC and a 194.6 GiB UD-IQ1_M quantization.
A measured recipe for the smallest NVIDIA box that holds the model: two DGX Sparks, a 194.6 GiB IQ1_M quant, llama.cpp RPC over the 200 GbE fabric, and roughly seven tokens per second.
- GitHub
GitHub
muse-spark-open-bench
An open benchmark that measures Muse Spark Open as a code reviewer on real merged PRs, including 25 changes that shipped a bug and 10 visual bugs with screenshots.
An independent code-review benchmark: 50 real merged pull requests, 25 of which were later reverted or fixed, run through Muse Spark Open with every catch checked by running the code.
Why the list is this short
An 816B model does not produce a wave of weekend projects the way a 421M one does. The repositories that exist are about reaching the model at all: a runtime recipe, a memory trick, a benchmark harness. That is the honest state of the ecosystem three days in, and it is more useful to say so than to pad the page with projects that merely mention the name.
Source: GitHub repository search Read 2026-10-17 Muse Spark Open topic