Meta releases Muse Spark Open, an 816B model that runs on two Mac Studios

Major

Lab claim

Meta published the weights for Muse Spark Open on October 15. It has 816 billion parameters, 42 billion active, and a 1 million token context. Meta says two M5 Ultra Mac Studios can hold it.

How to read this claim

Measurement — The full evaluation report is still to come. The desktop demos and kernel speedups are Meta's.
Bill — List input and output per million tokens on the vendor page — a loop pays both again.
Door — Status on this card: Shipped.

Keep the habit · field card with prices and worked checks

Two open agent models

Figures from each lab's release page, October 15 and 16, 2026.
	Beam	Muse Spark Open
Lab	Reflection	Meta
Date	October 16	October 15
License	Apache 2.0	Muse Model License
Total parameters	501 billion	816 billion
Active parameters	23 billion	42 billion
Context	1 million tokens	1 million tokens
Runs as described	Hosts listed by Reflection	Two M5 Ultra Mac Studios, or a host

What changed

Muse Wonder, released three days earlier, is the model you call. Spark Open is the weights you can hold. Before October 15 that agent-class model was not a file a team could download.

What it introduces

Meta's open agent model is now a file a team can hold, not only an API. The practical limit is the machine: two M5 Ultra Mac Studios, or a host.

Who benefits Teams that want open weights for coding and agent work. Anyone deciding between calling Muse Wonder and running a model they can tune.

Try this

If you need the weights, start from the Hugging Face collection and the license. If you need it today without two Mac Studios, use one of the listed API hosts and price that host separately.

What we do not know

The full evaluation report is still to come. The desktop demos and kernel speedups are Meta's.

Evidence

Primary Meta. Reported by Meta. The speedups and the desktop demos are Meta's.

Independent No independent rerun is cited. The two-Mac-Studio setup and the kernel speedups are Meta's.

Sources

Meta Primary · October 15, 2026

Still unclear? · Start here · ranks, prices, unknown The full evaluation report is still to come. The desktop demos and kernel speedups are Meta's.

The record

What happened

On October 15, 2026, Meta released the weights for Muse Spark Open. The model has 816 billion parameters in total and 42 billion active, with a 1 million token context, and it takes more than one kind of input. Meta says it is the first model in this class to run on a desktop, using two Apple M5 Ultra Mac Studios and MLX. Each Mac Studio can use up to 512 GB of unified memory, and the chip's memory bandwidth is 1.2 TB/s, so the pair holds about a terabyte. Weights are on Hugging Face and on Amazon, Google, and Microsoft model hosts. API access is listed at Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, and Together. The license is the Muse Model License, with limited commercial terms. Meta also says custom GPU kernels in SGLang deliver 1.03 to 3.95 times kernel-level speedups, and a 3 percent end-to-end speedup on AMD MI350.

Why it matters

Muse Wonder, released three days earlier, is the model you call. Muse Spark Open is the model you can download. A team that wants to fine-tune or inspect an agent model no longer has to stop at the API. The practical limit is the machine: Meta's own setup is two M5 Ultra Mac Studios, not a laptop. Hosted APIs are the path if you do not have that hardware. Meta says the model met the release thresholds in its Superintelligence Scaling Framework. The full evaluation report is still to come, so the safety sentence is Meta's for now.

Meta says it tested agent workflows entirely on the machines, including shipping iOS changes to the Ice Cubes app and driving DaVinci Resolve through a cut. Those are demonstrations, not a public score anyone else has rerun.

Further tuning is described through PyTorch and TorchTitan. Apple silicon uses MLX, which pulls the weights from Hugging Face.