Meta and Apple team up on an AI model built to run on Mac Studios instead of the cloud

Meta released a new open-weight artificial intelligence model this week that it says can run entirely on consumer Apple hardware rather than in the cloud, the result of an engineering partnership between two companies that have spent years competing rather than collaborating on AI. The model, called Muse Spark Open, was unveiled Oct. 15 in a post on Meta's official AI research blog.

Built by Meta's Superintelligence Labs, Muse Spark Open is a mixture-of-experts model with 816 billion total parameters, of which 42 billion are active for any given token, and a context window of 1 million tokens. What sets it apart from most large open-weight models is not its size but where it can run: Meta engineered the release in partnership with Apple so the model operates locally, with no cloud connection required, across two Apple M5 Ultra Mac Studios linked over Thunderbolt 5.

The numbers

The model's weights are distributed in both BF16 and a more compact NVFP4 format; even the smaller NVFP4 version requires roughly 493 gigabytes of memory, meaning it needs at least two Mac Studios configured with 512 gigabytes of unified memory apiece to run at all. According to benchmark data Apple published on its own machine-learning research site, four M5 Ultra machines running in tensor-parallel configuration produced roughly 36.7 tokens per second during decoding at an 8,000-token context length, compared with 15.7 to 25.7 tokens per second on two machines. Apple said its new M5 Ultra chip delivers up to 4.6 times faster time-to-first-token and 1.6 times faster generation than the previous M3 Ultra on other dense models it tested, using a new distributed-inference backend built on Apple's MLX framework and Thunderbolt-based remote direct memory access, which engineers have nicknamed "jaccl." Meta's own model card on Hugging Face confirms the parameter count, release date and licensing terms.

The model is also available through conventional cloud channels, including Amazon SageMaker, Google Cloud's Model Garden and Microsoft Foundry, and through inference providers such as Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together AI, which prices access at $1.25 per million input tokens and $4.25 per million output tokens. That roster gives developers who lack the hardware, or the roughly $10,000-plus combined cost of two qualifying Mac Studios, a way to use the same weights without buying into Apple's ecosystem at all. Meta's license, called the Muse Model License, is free for most developers but requires a separate commercial agreement once a company's API-resale revenue built on the model passes $20 million a year.

Context: a reversal after Meta's closed spring release

Muse Spark Open follows Meta's April release of a closed version of Muse Spark under Alexandr Wang, the former Scale AI chief executive who now leads Meta's Superintelligence Labs as its chief AI officer. That closed launch had fueled speculation that Meta was stepping back from the open-weight strategy that defined its earlier Llama model family. This release reverses that impression, putting a large, state-of-the-art model's weights back in developers' hands even as it ties the model's most distinctive capability — fully local, offline operation — to premium Apple hardware that, in the two-Mac-Studio configuration Apple benchmarked, costs well over $10,000 combined.

"Runs great on new Mac Studios thanks to work with [John Ternus] and Apple." — Meta chief executive Mark Zuckerberg, posting on X

Reaction

Zuckerberg credited Apple's hardware engineering chief, John Ternus, by name in the post; Ternus replied praising the efficiency of running "a model of this caliber and scale" from a desk rather than a data center. Wang, in a subsequent interview with Axios, joked about Meta and Apple "playing nice again" and credited Daniel Gross — a former Apple executive now at Meta — with helping broker the partnership, according to a report on the exchange between the two executives. Neither company disclosed financial terms, an exclusivity period, or the duration of the arrangement. Independent technology outlets including MacStories' technical breakdown of the release corroborated the specifications and the framing of the project as a genuine joint engineering effort rather than a simple compatibility test, within hours of the announcement.

Who is affected

The release matters most immediately to AI developers and enterprises that want large-model inference without sending data to a cloud provider — a priority for regulated industries such as health care, law and finance, where data residency and privacy rules can make cloud-based inference legally fraught. It is also, more directly, a sales pitch for Apple's priciest desktop hardware: the M5 Ultra Mac Studio configuration Apple used in its own benchmarks had not yet begun shipping as of the announcement, with availability expected later in October. For competitors in the open-weight race, including OpenAI, Google and xAI, the release raises the bar on what "open" now means, by tying a model's practical usability as much to a specific hardware ecosystem as to its weights being downloadable at all.

What happens next

Meta said a full evaluation report under its Superintelligence Scaling Framework — the internal safety-review process the company adopted for its most capable models — is "coming soon" but had not been published as of the weekend. That gap leaves independent researchers without Meta's own account of the model's risk profile even as the weights are already downloadable. Apple, for its part, has not said whether the distributed-inference software that makes two-Mac-Studio operation possible will be extended to other third-party models beyond Muse Spark Open, or whether the jaccl backend will become a standard feature of future MLX releases.