The qualifier is in the claim

Meta's headline is careful, and readers should be as careful as the headline. Muse Spark Open is presented as the most capable US open-weight model to date. That superlative is Meta's own positioning, not an independently established ranking, and the geographical word is not decoration. In the benchmark table on Meta's own model card, the model's 39 on the Artificial Analysis Intelligence Index sits ahead of the other US entries listed and ahead of Mistral Large 4 at 38, and behind Qwen3.8 at 40, Kimi K3 at 44 and GLM 5.3 at 45.

On agent measures the profile is mixed, even inside that table. Muse Spark Open tops MCP-Atlas at 89.2 against Kimi K3's 82.3, and edges JobBench at 61.8 against GLM 5.3's 61.4 and Kimi K3's 54.3. But it trails Kimi K3 on OSWorld Verified, 80.4 to 84.8, and on two coding benchmarks: Terminal-Bench 2.1 at 76.4 against 85.0, and DeepSWE 1.1 at 60 against 68. That is a coherent story — a very large open model tuned for tool use, multi-step coding and reading screenshots, documents and video while it works — with genuine strengths and visible gaps. It is not the same story as best open model, full stop, and Meta's own numbers are the source for both versions.

A desktop that is not on sale yet

The second headline is local running. Meta worked with Apple to optimise the model for the Mac Studio with M5 Ultra under Apple's MLX framework, and demonstrated agent workflows running entirely on device, including shipping changes to an open-source iOS app and driving a video editor. Two linked machines, each able to address up to 512GB of unified memory, can hold the model between them.

Read the conditions before treating this as a local-AI milestone you can buy, and keep the two deployment routes separate, because they are different paths rather than one claim contradicting the other. The model card's conventional GPU guidance asks for at least 700GB of VRAM for the NVFP4 checkpoint and 2TB for BF16, with the BF16 checkpoint itself about 1.5TB on disk. The Mac pair is a distinct MLX optimisation, not an exception to those figures. It is also not yet purchasable: the 512GB configuration in the demonstration has not been released; MacRumors reports it is due later this month, and the 256GB Mac Studio already costs over $10,000. The Mac result is a genuine engineering result and a narrow one. For almost every team reading this, Muse Spark Open is an API model that happens to have downloadable weights, not a desktop model.

Open weights, written conditions

The licence decides how much the download is worth. This is the Muse Model License, not Apache or MIT, and Hugging Face access is gated behind a form asking for a full legal name, date of birth and a full organisation name with corporate identifiers. Three terms deserve attention before a product team builds on it. Section 2 requires the licensee to prominently display "Built with Muse Spark Open" in the user interface of each product or service that uses the Software or any Derivatives, or in the accompanying documentation, website or settings menu — and the licence defines Derivatives broadly: any AI model created, trained, modified (including fine-tuned), distilled, or otherwise improved using the Software or any outputs or results thereof, so a model trained on Muse Spark Open's outputs carries the display duty too. Internal use, research and embedding the model inside your own product sit outside the revenue clause. But offering the model itself as a service is treated differently: once a licensee and its affiliates pass $20 million in aggregate gross revenue in any consecutive twelve months, they must notify Meta and enter a separate written agreement before running that model-as-a-service business. The usage policy also bars, among other things, military and critical-infrastructure operation.

None of that makes the release meaningless. Weights, a model card with full architecture detail, fine-tuning paths and seven API providers on day one are a real opening, especially after a period in which Meta's newest Muse models arrived without weights. It does mean the word open in the product name describes access to the weights, not the absence of a landlord.

What to do with it

Start with a hosted API and compare Muse Spark Open against the model you already pay for on representative work: a multi-file refactor, a long debugging session, a tool-using workflow you can score. The 1M-token context and the multimodal inputs are testable on your own material within an afternoon. And be clear-eyed about what the weights route involves: open weights here do not mean anonymous, frictionless downloading — the Hugging Face access form requests identifying and legal information before the files open. Leave the hardware decision alone until independent results exist and, for the Mac route, until Apple actually sells the machines in the demonstration. If your plan is to resell access to the model, read the $20 million clause before you price anything, not after.

Sources

  1. [1] Hugging Face model card, meta-models/Muse-Spark-Open-816B-A42B (primary model card — release date 15 October 2026, Muse Model License text including the Section 2 attribution requirement and the Section 3 $20 million model-as-a-service clause, 816B/42B parameters, 1M context, modalities, architecture, benchmark table, and the 700GB NVFP4 / 2TB BF16 self-hosting requirements plus the two-Mac-Studios MLX optimisation)Read source
  2. [2] Hugging Face, "Welcome Muse Spark Open" (launch post, 15 October 2026 — architecture detail, BF16 checkpoint about 1.5 TB, NVFP4 deployment, inference and fine-tuning paths)Read source
  3. [3] MacRumors, "Meta and Apple Bring Muse Spark Open AI Model to Mac Studio" (15 October 2026, 10:47 PDT — announcement timing, the two 512GB M5 Ultra Mac Studio MLX setup, the 512GB configuration not yet released and due later this month, the 256GB model priced over $10,000, API and download channels)Read source
  4. [4] 9to5Mac, "Meta launches new Muse Spark Open model, optimized to run on Mac Studio" (15 October 2026 — Meta's release wording, the on-device Muse Code workflows Meta says it tested, and the months-long Apple–Meta collaboration as described to Axios)Read source
  5. [5] Briefly, "Meta ships Muse Spark Open, its answer to China's open-weight models" (15 October 2026 — independent summary of the licence's revenue-notification, attribution and restricted-use terms)Read source