Meta launches Muse Spark Open, an open-weight AI model built for Mac Studio

Meta said the model is natively multimodal, with 816 billion total parameters, 42 billion active parameters and a 1 million-token context window. CNBC reported that it uses a mixture-of-experts architecture. According to Meta, Muse Spark Open is built for coding and agentic work and excels at long-horizon tasks, complex software engineering and multi-step reasoning.

The weights are available now through Hugging Face, and Meta is also making the model available through a number of API platforms, according to TechCrunch. Apple published a technical post about MLX, its open-source array framework for machine learning on Apple silicon, and about running Muse Spark Open and two other open models. Apple's Machine Learning Research website said MLX on Mac lets AI researchers and developers experiment locally with latest models and methods in a private environment.

Meta optimized Muse Spark Open with the latest version of MLX. The company said Apple's unified memory architecture lets the M5 Ultra's integrated GPU use up to 512 GB per Mac Studio, so two machines can hold an entire model. The chip's 1.2 TB/s of memory bandwidth is fast enough to read the small slice each token needs, Meta said, so two Mac Studios behave like one supercomputer with a terabyte of unified memory.

Running the model locally still requires substantial hardware. Engadget reported that Muse Spark Open needs at least 1 TB of RAM and will run best on Mac Studio machines equipped with the M5 Ultra chip. CNBC reported that a Mac Studio with the required M5 Ultra chip and 256 GB of memory costs at least $10,799, and that a more expensive 512 GB version will be released later this month. Engadget said Apple had not announced specific pricing or a ship date for the 512 GB Mac Studio, only that it would be available in late October. Engadget also said the model can run on four 256 GB M5 Ultra Mac Studios, a setup that would cost roughly $38,000.

Meta tested the model on a range of agentic workflows running entirely on-device using Muse Code, its AI coding terminal powered by Muse Spark Open. The company said it shipped real iOS changes to Ice Cubes, an open-source Mastodon client, including bug fixes, new features and a full build. It also drove DaVinci Resolve end to end, with the model reviewing footage frame by frame to find the strongest moments and crop them vertically.

Meta released evaluation results claiming superior performance to other popular open-weight models. Engadget reported that benchmarks released by Meta suggest Muse Spark Open outperforms two leading Chinese models, GLM 5.3 and Kimi K3, on agentic tasks and visual reasoning. Meta has said it believes the United States should not allow Chinese labs to lead on open-source AI. As an open-weight model released under a license allowing limited commercial use, Muse Spark Open can be downloaded and run on a wide range of hardware, CNBC reported.

Meta CEO Mark Zuckerberg publicly backed the launch. 'Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple,' he said on X. New Apple CEO John Ternus 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!' Meta AI chief Alexandr Wang told CNBC, '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.' Wang said many powerful open-weight models are too big to run on most personal computers, but Apple's Mac Studios make enough memory available to make running Muse Spark Open practical.

Representatives from Apple and Meta declined to say whether the collaboration was the start of a deeper relationship, CNBC reported. The two companies have a long history of rivalry. Apple introduced privacy changes in 2021 that limited digital ad targeting, and Meta estimated in early 2022 that the change would cut its revenue that year by $10 billion. The companies partnered on iPhone extensions more than a decade ago but have not struck any public partnerships since 2017, after Apple removed system-level Facebook integration from the iPhone. Tim Cook, who was Apple CEO until Ternus succeeded him last month, criticized Meta during the Cambridge Analytica scandal, and Zuckerberg told Joe Rogan last year that Apple 'hasn't invented anything great since the iPhone.'

Meta's AI strategy has also shifted. The company once centered its work on the older open-weight Llama family of models, but it has since focused on proprietary models under Wang, who joined Meta after the company invested more than $14 billion into his former startup Scale AI last year. CNBC reported that Meta bought hundreds of Mac computers earlier this year so its Superintelligence team could experiment with local AI and agents. The launch could challenge OpenAI and Anthropic, CNBC reported, and could signal Apple's growing strength in AI chips, a market Nvidia has dominated through large data centers.