What is new

Meta has released a new large language model called Muse Spark Open, describing it as the most capable open-weight model built in the United States to date. The announcement is notable less for the model's raw size than for who helped bring it to market: Apple. The two companies, which compete directly in consumer hardware and services, worked together so that Muse Spark Open can run on a pair of Apple M5 Ultra Mac Studios using Apple's MLX machine learning framework. Meta says this makes it the first model in its weight class that can run this way on desktop hardware.

The collaboration was made public with unusually warm words from both sides. Meta CEO Mark Zuckerberg credited Apple's John Ternus and his team for helping the model run well on the new Mac Studios. Ternus, in turn, praised the efficiency of running a model of this scale from a desk. Meta's chief AI officer, Alexandr Wang, told Axios the two companies had been working together for several months, joking that it was nice to see "mommy and daddy" cooperating again.

How it works

An open-weight model is one whose trained parameters, the numerical values that encode what the model has learned, are published for anyone to download and run, rather than being locked behind a company's own servers and API. This lets developers and researchers run the model on their own hardware, inspect it, and fine-tune it for specific tasks, though it does not necessarily mean the underlying training code or data are public too.

Muse Spark Open has 816 billion total parameters, a measure of the model's scale and, roughly, its capacity to store patterns learned during training. Meta describes it as natively multimodal, meaning it was built from the start to handle more than just text, rather than having other modalities such as images bolted on afterward. The company says it is particularly strong at long-horizon tasks (work that requires sustaining a plan over many steps), complex software engineering, and multi-step reasoning, where a model has to chain several logical steps together rather than answer in one shot.

Running a model this large normally requires data-center-grade hardware with large amounts of specialized memory. Apple's contribution is MLX, an open-source array framework designed specifically to make machine learning efficient on Apple Silicon chips, the processors Apple designs for its own Macs. MLX is built to take advantage of the way memory is shared between the CPU and GPU in Apple Silicon, which can make it practical to run very large models on consumer-facing hardware rather than specialized servers. Apple published a technical post alongside the launch covering MLX itself, Muse Spark Open, and two other open models, giving developers detail on how the framework was adapted to handle a model of this size.

The specific configuration Meta highlights, two Apple M5 Ultra Mac Studios working together, points to the model being split across two machines' worth of memory and compute, since a model with hundreds of billions of parameters typically needs more memory than a single consumer machine provides.

Why it matters

For practitioners, the headline is less about any single benchmark and more about access. If a model with hundreds of billions of parameters can genuinely run on desktop hardware rather than a cloud cluster, that lowers the barrier for researchers, startups, and hobbyists who want to experiment with a frontier-scale open model without renting expensive server time. It also signals that Apple is positioning Mac Studio and Apple Silicon generally as credible platforms for serious on-device AI work, not just a convenience for smaller models.

The fact that Meta and Apple worked together at all is itself a signal worth noting. The two companies are rivals across hardware, app stores, and AI assistants, and this kind of direct technical collaboration between them is unusual. It suggests both companies see enough mutual benefit, Meta in showcasing its model's efficiency, Apple in showcasing its hardware and MLX framework, to set aside that rivalry for this release.

  • Muse Spark Open: Meta's new natively multimodal model, 816 billion total parameters
  • Meta calls it the most capable open-weight model built in the US to date
  • It is described as the first model in its class able to run on two Apple M5 Ultra Mac Studios using Apple's MLX framework
  • Meta says it excels at long-horizon tasks, complex software engineering, and multi-step reasoning
  • Apple published a technical post covering MLX, Muse Spark Open, and two other open models
  • Meta says the two companies collaborated over several months before the joint announcement

Limits and open questions

The announcement does not give independent benchmark scores, so Meta's claim of being the most capable open-weight model to date cannot be verified from this material alone; it is Meta's own characterization. It is also not clear from what is described exactly how the model's workload is divided between the two Mac Studios, what speed or latency users should expect, or what the cost of that two-machine setup would be compared with cloud alternatives. The source does not specify licensing terms for Muse Spark Open's weights, nor does it describe the training data, safety evaluations, or how the model compares head-to-head against other open or closed models on standard reasoning or coding benchmarks. Readers interested in those details will need to look at Apple's technical post and Meta's own model documentation directly.

Sources