For most of the AI boom, running a frontier-class model meant renting time in someone else's data center. On Thursday, Meta showed a different picture: two Mac Studios on a desk, linked together, running one of the largest AI models anyone has released for download.
Meta has launched Muse Spark Open, an open-weight member of its Muse Spark family, and calls it "the most capable US open-weight model to date." The company worked with Apple to make it run locally on Mac Studio computers with the M5 Ultra chip, an unusual partnership between two companies that have clashed for years over privacy, app stores and advertising.
The weights are available now on Hugging Face and the major cloud platforms, so developers can download, run and fine-tune the model instead of only reaching it through Meta's servers.
Muse Spark Open has 816 billion parameters, 42 billion of them active at a time, and a 1 million token context window. Meta says two linked M5 Ultra Mac Studios can run it entirely on-device.
READ ALSO: Google's Gemini 4 Argon Is Out: Who Gets It and What It Costs
What Meta Released
According to Meta's announcement, Muse Spark Open is natively multimodal and built for coding and agentic work: long-running tasks, complex software engineering and multi-step reasoning. Only 42 billion of its 816 billion parameters are used on any single step, which is part of what makes it practical to run outside a data center.
Meta says the model beats other popular open-weight models on its own evaluations, but the full evaluation report is still listed as "available soon." Until independent testers publish their own numbers, treat the performance claims as Meta's.
The weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted access is available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, and the model works with common inference tools including vLLM, SGLang, llama.cpp and Apple's MLX.
It ships under the Muse Model License, which Meta says has limited commercial terms that "allow many developers to use the model for free." That is not the same as a fully open source license, so companies should read the terms before building a product on it.
Why the Mac Studio Matters
The headline feature is where the model can run. Each M5 Ultra Mac Studio can give its built-in GPU up to 512GB of unified memory, with 1.2TB per second of memory bandwidth. Linked together, Meta says, two of them "behave like one supercomputer with a terabyte of unified memory," enough to hold the whole model. TechCrunch, citing Engadget, notes that at least 1TB of memory is needed.
Meta used the latest version of MLX, Apple's open-source machine learning framework, to split the model across machines. Apple says MLX lets developers "experiment locally with the latest models and methods in a private environment," as quoted by 9to5Mac.
To show what this looks like in practice, Meta ran Muse Code, its AI coding terminal, entirely on-device. The model shipped bug fixes and new features to Ice Cubes, an open-source Mastodon app for iPhone, and drove DaVinci Resolve to find the best moments in raw footage and crop them for vertical video.
Meta's chief AI officer, Alexandr Wang, told Axios the two companies had worked together for several months, according to 9to5Mac. CNBC framed it as Apple and Meta teaming up to bring Muse to Mac computers. Neither company has said whether it signals a wider alliance.
READ ALSO: Anthropic Warns AI Could Soon Copy Itself Onto Cheap Smart Devices
Meta's Open Model Strategy, Revisited
The release also settles a question that has hung over Meta since April. When it launched the original Muse Spark on 8 April, the first model from the team led by Wang, Meta kept the weights closed. That was a sharp turn for the company whose Llama models had made it the best-known American open-weight developer, although it said at the time that it planned to open future versions.
Muse Spark Open delivers on that promise at a scale few rivals match. Meta says the model received extra safety training aimed at the risks of open weights, and that it "met the thresholds our framework requires for release" under the company's internal safety framework.
What This Means for You
For developers and startups, a frontier-class model that runs on hardware you own changes the economics. Once it is running there are no per-token fees, and sensitive code or data never has to leave the building. The catch is the upfront cost: two top-spec Mac Studios are a serious purchase, and the license still carries commercial limits.
For businesses in regulated fields such as healthcare, finance and law, private AI that runs on premises is exactly what compliance teams have been asking for. Expect pilots before the end of the year.
For everyone else, this is a sign of where AI is heading. Capable models are moving from distant server farms toward machines that sit on a desk, and two of the biggest names in tech now share an interest in making that happen.
Muse Spark Open puts a frontier-scale AI model on two Mac Studios. If its performance holds up in independent tests, running serious AI privately on your own hardware just became a realistic option.
