Apple and Meta Bury the Hatchet to Put an 816B-Parameter Model on Your Desk
Meta released the weights of Muse Spark Open on October 15, and announced it had worked with Apple so the model can run locally on Mac Studio desktops instead of on remote servers. Apple's contribution is distributed inference through MLX, its array framework for Apple silicon: two or four Mac Studios with M5 Ultra chips are linked over Thunderbolt 5 and remote direct memory access, pooling their compute. According to Meta's launch post, the M5 Ultra's integrated GPU can address up to 512GB of unified memory per machine, so two paired machines behave like one supercomputer with a terabyte of unified memory and 1.2 TB/s of memory bandwidth.
Muse Spark Open is a sparse mixture-of-experts model: 816 billion total parameters, roughly 42 billion activated per token, a 1-million-token context window, natively multimodal, and built on the same base model as the closed Muse Spark 1.3. Apple's machine learning research post puts the 816B model at about 493GB even at 4-bit precision, and measured four M5 Ultra Mac Studios generating over 36 tokens per second at an 8,192-token context.
Meta's framing is blunt: the "most capable US open-weight model to date." Meta AI chief Alexandr Wang said in an interview that many powerful open-weight models are simply too big to run on the majority of personal computers, while Mac Studios make enough memory available to make local running practical. Meta's own evaluations put it ahead of Nemotron 3 Ultra, Mistral Large 4, Qwen3.8, GLM 5.3 and Kimi K3 on agentic tests such as MCP-Atlas and JobBench and on reasoning tests including HLE-Diamond and AA-Omniscience, while on some coding benchmarks it trails Kimi K3 and GLM 5.3.
" Muse Spark Open — weights releasing today — runs great on new Mac Studios thanks to work with John Ternus and Apple," Zuckerberg wrote on X. Apple's new chief executive replied: "Congrats to the team! Amazing to run a model of Muse Spark Open's caliber and scale this efficiently, right from your desk. Mac Studio keeps pushing the boundaries of on-device AI!" Representatives for both companies declined to say whether the collaboration is the start of something deeper.
What shipped is not Meta's best model. Three days earlier, on October 12, Meta launched the proprietary Muse Wonder and Muse Wonder Full, calling them its most capable yet. On independent evaluator Artificial Analysis's intelligence index, Wonder Full scored 56 and Wonder 52, against 48 for Spark 1.3; Wonder Full sits just behind Claude Opus 5.5. Bloomberg read the release as Meta giving developers something to download and modify while keeping its most advanced technology closed, in an effort to turn its AI spending into revenue. Wang conceded Muse Spark Open is less advanced than Muse Wonder — but good enough, he said, to entice more coders to upgrade.
"We're excited for mommy and daddy to be playing nice again," Wang told Axios, adding that the two companies had been working together for several months and crediting Daniel Gross, a former Apple executive now at Meta, with a key role in bringing them together. Apple published a technical post on MLX distributed inference the same day, using Muse Spark Open and two other open models as its worked examples.
Cost is the catch. A Mac Studio with the required M5 Ultra chip and 256GB of memory starts at $10,799, according to CNBC; a 512GB version arrives later this month. Engadget notes that running the model across four 256GB M5 Ultra Mac Studios costs roughly $38,000. The Wall Street Journal put it more starkly: buyers need Macs with at least 1TB of memory combined, at a total cost of more than $40,000. Apple has not priced or dated the 512GB configuration beyond "late October."
The weights are free to download from Hugging Face, but not under a standard open-source licence: Meta distributes them under its own Muse Model License, which permits many uses without charge while placing limits on commercial use. Distribution is broad — API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI offer the model, with weights also available via Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Meta shipped supporting optimizations too: custom Helion GPU kernels in SGLang deliver 1.03x to 3.95x kernel-level speedups on Nvidia and AMD GPUs.
Meta says the model also runs on AMD, Intel and Nvidia hardware and is optimized for the SGLang and vLLM inference engines; developers without the equipment can use hosted providers instead. Meta paired the model with Muse Code, its AI coding terminal, and showed an agentic workflow fixing up an open-source Mastodon client with bug fixes, new features and a full build.
Two things matter from here: Apple's pricing and ship date for the 512GB M5 Ultra Mac Studio, still promised only for late October, and whether the pairing grows into something more permanent — a question both companies declined to answer. The release also fits a pattern: a small open-weight model, Muse Glimmer, in August; the closed frontier model Muse Wonder on October 12; open weights on October 15. For developers, the practical meaning may be simpler: running a frontier-class model on your own desk now has a shopping list — and the list asks for two, or four, Mac Studios.