MENLO PARK, California — Meta released Muse Spark Open on Thursday, an 816-billion-parameter artificial intelligence model that anyone can download, and said it had worked with Apple so the model runs on two linked Mac Studio desktop computers.
The release brings a model of a size usually run on data-centre servers onto desktop hardware, albeit expensive hardware. It also joins two companies whose leaders have traded harsh words over privacy and advertising for years.
Meta calls Muse Spark Open "the most capable US open-weight model to date." Open-weight means the company publishes the model's parameters, the numerical values learned in training, so developers can run and modify it on their own machines instead of renting access through a cloud service.
The model uses a mixture-of-experts design, in which only part of the network works on each token, the word fragments AI systems read and write. Only 42 billion of its 816 billion parameters are active at a time, according to Meta's model card. It accepts text, images, video and audio, handles up to 1 million tokens of context and is built for coding and agentic tasks, in which software carries out multi-step jobs for a user.
Apple published its own test results the same day, saying a compressed 4-bit version needs about 493 gigabytes of memory, too much for one Mac once working memory is added. Two M5 Ultra Mac Studios linked by Thunderbolt 5 and running Apple's MLX software produced up to about 25 tokens a second. "On a cluster of four Mac Studios, Muse Spark Open generates over 36 tokens per second," Apple wrote.
Meta said the paired machines "behave like one supercomputer with a terabyte of unified memory." In company demonstrations, the model fixed bugs and added features in Ice Cubes, an open-source iPhone app for the Mastodon social network, and reviewed and cropped footage in DaVinci Resolve video software.
Mark Zuckerberg, Meta's chief executive, wrote on X that the model "runs great on new Mac Studios thanks to work with @johnternus and Apple." John Ternus, who succeeded Tim Cook as Apple's chief executive last month, replied: "Amazing to run a model of Muse Spark Open's caliber and scale this efficiently, right from your desk."
"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," Alexandr Wang, Meta's AI chief, told CNBC.
The two companies have struck no public partnerships since 2017, when Apple removed system-level Facebook integration from the iPhone, CNBC reported. Apple's 2021 privacy changes limited ad targeting, and Meta estimated in early 2022 they would cut that year's revenue by $10 billion.
"Most of the animosity predates my tenure at the company, but obviously there's been a long history between Meta and Apple," Wang said. He named Daniel Gross, an AI researcher and former Apple director who joined Meta last year, as a reason for the collaboration. Representatives of both companies declined to say whether the work would lead to a deeper relationship.
The hardware is costly. A Mac Studio with an M5 Ultra and 256 gigabytes of memory starts at $10,799, and the 512-gigabyte version Meta's two-machine setup relies on has no announced price. Engadget and VentureBeat estimated that four 256-gigabyte machines would cost roughly $38,000 to $43,000.
"For all but a small subset of researchers, you'd be much better off paying to run it on a cloud server," Jason Cross, a senior editor at Macworld, wrote. Meta lists Together AI, Groq and OpenRouter among services offering hosted access.
Meta's own benchmarks are mixed. Muse Spark Open scored 89.2 on MCP-Atlas, a tool-use test, ahead of the Chinese models GLM 5.3 at 84.2 and Kimi K3 at 82.3. It trailed both on Terminal-Bench 2.1, a coding-agent test, scoring 76.4 against Kimi K3's 85.0. On the Artificial Analysis Intelligence Index, a composite score, Meta lists its model at 39, below GLM 5.3 at 45 and Kimi K3 at 44. VentureBeat said it could not verify an independent Artificial Analysis assessment of the model.
That rivalry is the political backdrop. Chinese models accounted for 41% of downloads in the year to March, ahead of American ones, according to a Hugging Face assessment cited by VentureBeat. Meta says the United States should not let Chinese labs lead open-source AI, Engadget reported.
Nathan Lambert, who writes the Interconnects newsletter on AI, welcomed the release on X. "I waited so long for US open weight models to be relevant again," he wrote.
Wang said the open model sits below Meta's best. "These aren't the most capable models that we've developed or released," he told CNBC. Meta launched Muse Wonder, a proprietary model whose weights it has not released, on Oct. 12.
The licence carries conditions. Companies that sell the model as a hosted service and have more than $20 million in annual revenue must sign a separate agreement with Meta, and products built on it must display the phrase "Built with Muse Spark Open," according to the licence. Meta said the open version received extra safety training and met its internal risk thresholds, but cautioned that modifications could degrade its safety.
Meta said a full safety evaluation report would be published soon. Apple has said the 512-gigabyte Mac Studio will be available in late October.
Sources
- Apple and Meta team up to make Muse AI model run on Mac computers (CNBC) (cnbc.com)
- Meta launches an open-weight AI model that runs on Macs (TechCrunch) (techcrunch.com)
- Meta unveils Muse Spark Open, the new top U.S. open weights model, and it runs on two paired Mac desktops (VentureBeat) (venturebeat.com)
- Meta built a huge open-weight AI model for Macs, but you'll need 1TB of RAM to run it (Engadget) (engadget.com)
- No, Meta's Muse will not run on your Mac (Macworld) (macworld.com)
- Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio (Apple Machine Learning Research) (machinelearning.apple.com)
- Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date (Meta) (research.meta.ai)
- Muse-Spark-Open-816B-A42B model card (Hugging Face) (huggingface.co)
This report was compiled from the publicly available reporting listed above. Direct quotations are reproduced as published by those outlets. Corrections: see our corrections policy.