On Thursday local time, Meta officially released Muse Spark Open, the most capable open-weight model in its history to date, and announced a partnership with Apple to enable the model to run locally on Mac computers.
As a natively multimodal large model built on a mixture-of-experts architecture, Muse Spark Open has a total of 816 billion parameters (of which 42 billion are active) and supports a context window of up to 1 million tokens. Meta said the model is designed for coding and intelligent agent tasks, delivering outstanding performance on long-horizon task handling, complex software engineering problems, and multi-step reasoning. In benchmarks published by Meta, the model ranked first across six tests—workflow, tool calling, knowledge reasoning, knowledge reliability, complex visual reasoning, and image-based factual Q&A—and the company officially called it the most capable American open-weight model to date. However, the test data also showed that Chinese open-source models such as Kimi, Qwen, and Zhipu still outperform it on software development and certain other tasks.
One of the most closely watched highlights of this release is Meta's unexpected joining of hands with Apple. During Tim Cook's tenure leading Apple, the two companies have clashed many times—Cook accused Meta of violating user privacy, while Zuckerberg mocked Apple for lacking innovation. But driven by their shared interest in open-source AI, the two sides chose to bury the hatchet. Daniel Gross, co-lead of Meta's compute division and a former Apple director, became one of the key figures behind the partnership.
On the hardware and software adaptation side, Meta deeply optimized Muse Spark Open using the latest version of MLX, an open-source machine learning framework optimized for Apple silicon, making it a perfect fit for the Mac Studio equipped with the M5Ultra processor. Meta said the combination of the two is the best solution for running powerful models on a personal computer. According to official information, Apple's latest Mac Studio (M5Ultra) supports up to 512GB of unified memory, and by connecting two machines the model can run in full.
However, the high-end hardware configuration also brings a considerable cost barrier. Although Zuckerberg has already posted photos on social media of the model running on two fully specced Mac Studios, the 512GB version set by Apple is scheduled to go on sale at the end of month 10 (October). Based on the currently sold 256GB Mac Studio (priced at over 10,000 US dollars), building a cluster with roughly 1TB of memory would cost more than 40,000 US dollars. Despite the hefty price, for developers who have long needed to rent cloud compute, purchasing local hardware could pay for itself through saved rental fees, while also offering compute headroom and data privacy.
Meta AI head Wang Tao pointed out that although the two companies still compete in certain areas, that does not prevent them from cooperating in areas where it makes sense. In addition to Apple silicon, Muse Spark Open has also been comprehensively optimized for chips from Nvidia, AMD, and Intel, and is compatible with a variety of mainstream inference engines. Developers can call the API through multiple platforms, or choose to download the model weights and deploy it themselves.