Muse Spark Open: Meta’s 816B MoE Runs on Two Mac Studios

Muse Spark Open is Meta’s October 15 open-weight release: a natively multimodal mixture-of-experts model with 816 billion total parameters, about 42 billion active per token, and a 1-million-token context window, positioned for coding and long-horizon agent work.

What actually happened

Meta’s research blog frames Muse Spark Open as the first model in its class that it can run at home on a pair of M5 Ultra Mac Studios. Apple’s unified memory (up to 512 GB per Studio) plus 1.2 TB/s bandwidth is the hardware story: two machines behave like one terabyte-class box for a sparse MoE that only needs a slice of weights per token. Meta says it tested on-device agentic workflows with Muse Code, including iOS changes to the Ice Cubes Mastodon client and end-to-end DaVinci Resolve editing passes.

Availability is broad on day one. Inference partners listed on the blog include Together AI, Fireworks AI, Baseten, Modal, Groq, Nebius and OpenRouter. Weights are called out for Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models. Runtime support spans SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp and Unsloth, with Helion GPU kernels in SGLang claiming 1.03–3.95× kernel-level speedups on NVIDIA and AMD and about 3% end-to-end on AMD MI350.

Safety, license, and why this open-weight release matters

Meta says The model received additional safety post-training for open-weight release and that it met thresholds under the Meta Superintelligence Scaling Framework; detailed numbers live in the model card, with a fuller eval report promised later. The weights ship under the Muse Model License with limited commercial terms that Meta says allow many developers to use the model for free. Read the license — “many” is not “all.”

For the open-weight timeline, this sits next to other agent-stack news we have been tracking, including Shopify Shop MCP on the commerce-tooling side and Anthropic Scion on the internal-infra side. The release is the rare post that tries to sell both frontier-ish capability and a literal two-Mac-Studio home rack.

Open weights move the bottleneck from API keys to power circuits and license footnotes.

Hugging Face’s companion blog restates the multimodal MoE framing and points developers at the collection link Meta published. We are not reprinting benchmark tables Meta deferred to a separate methodology report.

Local silicon versus hosted endpoints

Meta’s two-Mac-Studio demo is a positioning play as much as an engineering flex: personal superintelligence rhetoric meets a shopping list. Unified memory architecture is the enabling trick for sparse models that would otherwise demand datacenter racks. Developers who never touch Apple silicon still get day-0 paths through vLLM, SGLang and the listed API hosts — which is likely how most memcool readers will actually poke the model.

Customization pointers on the blog run through PyTorch TorchTitan and Helion kernels. We are not reproducing kernel benchmarks beyond Meta’s stated ranges. If you need numbers for a procurement memo, wait for the evaluation methodology report Meta flagged rather than scraping a leaderboard screenshot.

Open-weight releases also change the threat model. Meta acknowledges that and cites extra safety post-training plus framework thresholds. That is the lab’s claim; third-party red teams will test it. Until those write-ups exist, treat “met thresholds” as a release gate statement, not an independent audit.

For agent builders comparing stacks, the multimodality claim (native, not bolted-on) is the differentiator Meta wants you to notice next to the parameter count. A 1M-token window matters only if your tooling can feed and retrieve from it without melting latency budgets. Start with a hosted endpoint for smoke tests, then decide whether local MLX is a product requirement or a demo flex. Our personal agent protocol notes are a useful adjacent reading list for agent interface design even though they predate this release.

What to do as a reader

If you are evaluating this Meta release, decide your path first: hosted API, cloud weights, or local MLX on Apple silicon. Local means budgeting for two M5 Ultra Studios and reading Meta’s MLX adapter notes; hosted means checking which partner snapshot you actually hit.

Read the Muse Model License before any commercial plan, skim the model card safety section, and treat “most capable US open-weight” as Meta’s claim pending the eval report. Nothing here is investment advice.

Primary sources: Meta’s Muse Spark Open blog and the Hugging Face announcement.