Muse Spark Open: an 816B Open Model on Two Mac Studios — Did It Beat China's Open Models?
Meta opened the weights again. Muse Spark Open, released on October 15, is an 816-billion-parameter open-weight model — and the stage prop is unusual: not a data center, but two Mac Studios on a desk. Meta calls it the most capable US open-weight model to date; the independent index says something else. Here is both, kept separate. Ribbit.
The short version
The release — 816B total, 42B active per token, 1M context, free weight download
The stage — Two M5 Ultra Mac Studios (about 1TB of unified memory), not a data center
The blank — "Best in the US" checks out. On the independent index, seven Chinese models still sit above it
What actually shipped
- Meta Superintelligence Labs released the Muse Spark Open weights on October 15. The official blog post and the Hugging Face model card are the primary sources.
- It is a mixture-of-experts model: 42B of the 816B parameters activate per token, choosing 8 of 256 experts. Context runs to 1 million tokens, and it natively takes text, image, video and audio in, and returns text. Reasoning depth is set with reasoning_effort.
- The targets are coding and long-horizon agent work. Meta says the open version got extra safety post-training so it holds up standalone, without system-level guardrails.
- Weights are downloadable from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry. Hosted APIs are live at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
Sources: Meta official blog · Reuters
Two Mac Studios — why that is the headline
- Working with Apple, Meta linked two M5 Ultra Mac Studios (512GB of unified memory each) over MLX — about 1TB combined — and ran the NVFP4 build locally. "Two machines behave like one supercomputer" is Meta's pitch.
- One condition: the 512GB Mac Studio ships later this month and its price is not announced yet. The 256GB M5 Ultra model on sale today already costs $10,799 — over ten thousand dollars. "Runs on your desk" starts after that hardware reaches the market.
- On GPUs, the card calls for at least 700GB of memory for NVFP4 and 2TB for BF16. Local runs are an equipment game, not a hobby download.
Side by side
| Measure | Muse Spark Open | Chinese open models |
|---|---|---|
| MCP-Atlas (agentic tool use) | 89.2 — top of the table | Qwen3.8 84.5 · GLM 5.3 84.2 · Kimi K3 82.3 |
| Terminal-Bench 2.1 (coding) | 76.4 | Kimi K3 85.0 · GLM 5.3 83.9 · Qwen3.8 82.0 |
| AA Intelligence Index (independent) | 39 — 8th among open weights | MiMo-V2.6-Pro 46 · GLM-5.3 45 · Kimi K3 44 |
| Weights & license | Downloadable (Muse Model License, conditional commercial use) | Varies by model; many large downloadable models |
| One line | The top US open-weight model | The open-weight summit is still Chinese |
Scores are not blended: MCP-Atlas and Terminal-Bench come from Meta's published comparison table; the AA Index is from independent evaluator Artificial Analysis.
What Meta showed it doing
- Paired with the Muse Code agent, it shipped iOS changes to Ice Cubes, an open-source Mastodon client — bug fixes, new features and a full build.
- It drove DaVinci Resolve to review footage frame by frame, pick the strongest moments and crop them into vertical video.
- Both are Meta's own demos. The positioning reads clearly: less a model that answers, more an agent that picks up tools and finishes the job.
Meta's week — a closed model and an open one
- Three days earlier, on October 12, came the closed frontier pair Muse Wonder and Wonder Full. Only the weaker Wonder is on the API ($4 in / $20 out per million tokens); the stronger Full lives inside the Muse Code subscription — a split VentureBeat flagged on day one.
- In August, Meta shipped the small open model Muse Glimmer, and Zuckerberg's essay that month criticized the closed labs: "the future belongs to everyone."
- Closed on Monday, open on Thursday. Spark Open is the flagship of the open side of that two-track week.
Sources: VentureBeat · Chosun Biz
The cold water
- Who made the table decides the winner. Even in Meta's own table, Chinese models lead on coding, and on the independent index seven Chinese models sit above Spark Open.
- The star of the "runs on your desk" story — the 512GB Mac Studio — is not out yet and has no price. Until it does, the economics of local runs cannot be calculated.
- The license requires a "Built with Muse Spark Open" label, and companies over $20M in annual revenue need a separate agreement with Meta to sell it as an API service. It is not plain open source.
- The weights are free; hosted runs are not cheap. Artificial Analysis lists hosted pricing at $1.25 in / $4.25 out per million tokens — expensive for its open-weight class.
Puzz, in one line
The news in Spark Open is not the number 816B — it is the direction. Meta is back to big open weights, and the stage is local hardware, not the cloud. The crown, though, still sits with China's open models. "Best in the US" is the accurate address. Ribbit.