Apple and Meta seem to team up to run Muse Spark Open on Mac Studios

Apple and Meta, longtime rivals in privacy and ads, appear to have found common ground on Thursday, October 15, around an open-weight model large enough to need a desk full of high-end Macs, according to Meta’s research blog, Apple Machine Learning Research, CNBC and The Verge. Meta released Muse Spark Open; Apple published fresh MLX benchmarks showing the same model running across linked Mac Studios.

Meta describes Muse Spark Open as a natively multimodal mixture-of-experts model with 816 billion total parameters and 42 billion active, plus a 1-million-token context window, and says it built the release for coding and agentic work. Meta calls it “the most capable US open-weight model to date” and “the first model in its class” to run on a desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework. This desk could not independently re-run Meta’s capability rankings.

Apple’s MLX post, also dated October 15, walks through how the open-source array framework now splits frontier-class models across Macs over RDMA on Thunderbolt 5. Apple says macOS 26.2 introduced that low-latency RDMA driver, and that MLX version 0.30.1 and later add communication collectives branded jaccl. On four Mac Studios, Apple says Muse Spark Open generates over 36 tokens per second at an 8,192-token context — “well above average human reading speed,” in Apple’s words.

Memory sets the floor. Apple says Muse Spark Open would need more than 1.6 TB of RAM in native BF16 for the parameters alone, but an NVFP4 (4-bit) checkpoint cuts that footprint to about 493 GB, which still will not fit on one Mac Studio once intermediate compute needs enter the picture. Meta’s post says Apple’s unified memory lets each M5 Ultra Mac Studio use up to 512 GB, so two machines can hold the model, and cites 1.2 TB/s of memory bandwidth per chip. Apple’s cookbook-style numbers match the two-machine and four-machine layouts Meta’s hardware docs describe.

The companies’ leaders put the collaboration on X. The Verge and CNBC quote Meta CEO Mark Zuckerberg: “Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple.” Apple CEO John Ternus replied, per both outlets: “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!” The Verge also quotes Meta chief AI officer Alexandr Wang telling Axios, “We’re excited for mommy and daddy to be playing nice again,” and notes months of collaboration.

CNBC places the work in a longer rivalry. It says Tim Cook and Mark Zuckerberg clashed for years over privacy and ads, that Apple’s 2021 tracking changes led Meta to estimate a $10 billion revenue hit in 2022, and that John Ternus succeeded Cook as Apple CEO last month. CNBC quotes Wang: “We worked with Apple to bring leading Agent capabilities to your desktop for the first time on the Apple M5 Ultra Mac Studios, so you know we’re really excited about this.” Representatives from both companies declined to say whether Muse Spark Open marks the start of a deeper relationship, CNBC reports.

Cost remains a barrier for hobbyists. CNBC says a Mac Studio with the required M5 Ultra chip and 256 GB of memory costs at least $10,799, that a more expensive 512 GB version arrives later this month, and that buyers still need multiple machines. Meta’s blog frames the same hardware as “the best way to run a powerful model on your own desktop.” Neither outlet claims a single consumer laptop can hold Muse Spark Open.

Meta says developers can pull weights from Hugging Face and run them with MLX on Apple silicon, or serve the model through SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp and Unsloth on AMD, Apple, Intel and NVIDIA hardware. It lists hosted inference partners Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Meta says it releases the weights under the Muse Model License with limited commercial terms that “allow many developers to use the model for free.”

Meta also says Muse Spark Open carries additional safety post-training for open-weight release and that evaluations under its Meta Superintelligence Scaling Framework met the release thresholds; a full evaluation report, Meta says, arrives soon. This desk did not read the model card’s full safety tables. Apple’s post focuses on speed, power and parallelization rather than safety claims.

Meta’s on-device demos, described in its blog, include shipping iOS changes to Ice Cubes, an open-source Mastodon client, and driving DaVinci Resolve end to end with frame-by-frame review. Apple’s post shows how to launch an OpenAI-compatible local server with mlx_lm.server across a Thunderbolt mesh. Both accounts present local agentic work as the pitch, not a cloud replacement for every workload.

Reading note: this desk wrote from company posts and published reports, not from a Mac Studio cluster. Capability rankings, token-per-second figures and safety thresholds come from Meta or Apple; later independent benchmarks may differ.