Meta and Apple Bring 816-Billion-Parameter Muse Spark Open Model to Mac Studios

Meta and Apple are working together to let developers run Meta’s Muse Spark Open model locally on high-end Mac computers, with clustered Mac Studio systems providing the hardware base. The collaboration links Meta’s open-weight model strategy with Apple’s silicon and desktop infrastructure, according to company announcements and coverage from CNBC, The Verge, and TechCrunch.

Muse Spark Open has approximately 816 billion parameters, with about 42 billion active for each token, and supports text, images, audio, and video. Its context window reaches 1 million tokens, giving developers a model designed for large documents, multimodal analysis, and complex agent workflows. Meta describes it as the most capable open-weight model released by a U.S. company to date.

The arrangement turns Mac Studio clusters into a local platform for a model that would normally demand substantial data-center capacity. Apple silicon’s unified memory makes large-model inference more practical on a desktop, while local execution gives developers a way to test and operate AI systems without sending every prompt and file to a remote provider.

The hardware requirement remains significant. Muse Spark Open is not a lightweight laptop model like Meta’s 30-billion-parameter Muse Glimmer, which was built for local agents on ordinary computers. Instead, Spark targets developers with access to multiple high-memory Macs, placing it between consumer on-device AI and centralized cloud inference. Newsquawk described the partnership as a familiar AI hardware-cycle exchange: Meta gains distribution for its model, while Apple gains a credible local-computing demonstration without developing the model itself.

Muse Spark Open expands Meta’s effort to compete with OpenAI, Anthropic, and Chinese open-model developers through openly available weights and broad developer access. The model follows Muse Glimmer, which Meta positioned as an open-weight system for local agents, and arrives alongside other Muse-branded releases aimed at personal and professional agent tasks.

The company’s strategy combines frontier-scale research with models that developers can download, adapt, and run outside Meta’s own services. Mark Zuckerberg has also argued for American open-source AI and criticized restrictions that he says benefit foreign laboratories. That policy message gives the Mac collaboration a second purpose: it presents local, accessible computing as part of the U.S. response to intensifying international competition.

Developers gain a local environment for experimenting with a multimodal, long-context model without depending entirely on cloud APIs. Local inference can help with privacy-sensitive files, offline testing, latency control, and applications that need sustained access to large amounts of context. It also gives researchers more control over model behavior and deployment choices.

The tradeoff is cost and complexity. Running an 816-billion-parameter mixture-of-experts model requires substantial memory, multiple systems, and careful software configuration. Mac-based deployment therefore serves studios, research teams, and well-funded developers more directly than typical consumers. Meta’s separate Muse Glimmer release remains the more practical route for laptop-scale local agents, while the company’s Muse consumer product focuses on personal tasks across Mac and mobile devices.

Apple and Meta have spent years clashing over privacy, advertising, platform control, and the future of computing. Their cooperation on Muse Spark Open stands out because it joins a company known for tightly controlled hardware and software with a company using open model weights to widen distribution.

The alliance is commercially practical. Meta needs capable local hardware and a strong developer story for its models. Apple benefits when Mac systems become associated with serious AI workloads, particularly as developers evaluate alternatives to cloud-based inference. The relationship does not erase their broader competition, but it creates shared ground around memory-rich desktop computing and private AI processing.

The immediate test is whether developers can make Mac Studio clusters useful enough to justify their expense and operational demands. Model availability alone won't guarantee adoption: tooling, compatibility, inference speed, memory management, and documentation will determine whether Spark becomes a working platform or a technical showcase.

Meta’s expanding Muse family gives developers several deployment targets, from laptop-scale agents to frontier multimodal systems. Apple’s role places Mac hardware inside that progression and strengthens the case for local AI development. If the collaboration produces reliable tools and competitive performance, it will add pressure on cloud providers and hardware rivals to make large-model inference easier outside centralized data centers.