On October 15, 2026, Meta released the weights of its new open model, Muse Spark Open, after optimizing it in collaboration with Apple to run locally on Mac Studio devices powered by the M5 Ultra chip. This collaboration marks the first public partnership between the competing giants since 2017 to provide a strong alternative to Nvidia processors and closed models.
Launch of the Muse Spark Open model with multimodal capabilities and 816 billion parameters
High hardware cost and historical truce following a estrangement since 2017
Strategic alternative to Nvidia processors and groundwork for adopting the Muse Wonder model
Frequently asked questions
Launch of the Muse Spark Open model with multimodal capabilities and 816 billion parameters
Meta Platforms released the software weights of its new model, Muse Spark Open, on Thursday, October 15, 2026, describing it as “the most efficient and capable open-weight American model to date.” The company announced that it worked jointly with Apple to tune and optimize the model to run locally with high efficiency on the Mac Studio computer equipped with the M5 Ultra chip, according to 9to5Mac, in a deal that brought together two companies that remained in fierce competition for years, described by Axios as “a striking breakthrough in relations.”
The Muse Spark Open model handles multiple types of inputs natively and directly, featuring a total of 816 billion software parameters, with 42 billion active at any given time, alongside a massive context window of approximately 1 million tokens. Meta built this model specifically for coding tasks and intelligent agent operations, running on two Mac Studio devices or 4 devices connected together using Apple’s MLX framework.
The model is available for free download with a license permitting commercial use, according to Axios, while CNBC reported that Meta describes the license as allowing limited commercial use. Meta’s standard benchmark tests indicate that the model outperforms the Chinese models GLM 5.3 and Kimi K3 in agentic and visual reasoning tasks, according to Engadget.
High hardware cost and historical truce following a estrangement since 2017
Despite the software being free, the hardware required to run it is prohibitively expensive; Engadget notes that running the model requires about 1 terabyte of RAM. Meta explains that each Mac Studio device can allocate up to 512 gigabytes of memory to the GPU, allowing “two Mac Studio devices to work together as a single supercomputer with a full terabyte of unified memory.” Apple has not yet announced the price of the 512GB version scheduled for release in late October, while a setup consisting of 4 devices with 256GB capacity costs about $38,000 USD.
Alexander Wang, head of artificial intelligence at Meta, told Axios regarding this rapprochement: “We are excited to see mom and dad playing nicely together again.” Wang revealed that Daniel Gross, a former Apple executive currently working at Meta, helped bridge the gap between the two companies, whose relationship severely deteriorated in 2021 when privacy changes at Apple restricted ad targeting, which Meta estimated reduced its 2022 revenue by about $10 billion.
CNBC pointed out that the two companies have not announced any public partnership since 2017, and this move comes shortly after John Ternus assumed the position of CEO of Apple, succeeding Tim Cook. Given that the model runs entirely on users’ private devices, Meta never views their prompts, with Wang emphasizing: “Meta has zero control or visibility into any of that whatsoever.”
Strategic alternative to Nvidia processors and groundwork for adopting the Muse Wonder model
According to an analysis by Axios, running a model of this massive scale on Apple hardware could provide businesses with a practical alternative to Nvidia’s graphics processors and proprietary closed models. Concurrently with the announcement, Apple published new performance benchmarks for its chips and technical documentation explaining how the MLX framework splits massive models across multiple devices, though neither party disclosed whether the partnership would go further.
Alexander Wang explained that the open model ranks just behind Meta’s proprietary closed Muse Wonder model, which was released earlier this week.
Wang told CNBC: “These are not the most efficient and capable models we have developed or launched; but we believe there will be plenty of companies and developers who will use this model and experiment with it on a small scale, and later seek to expand and scale up their usage.”
Frequently asked questions
Question: What are the technical specifications of Meta’s Muse Spark Open model?
Answer: The model features 816 billion total parameters with 42 billion active at a time, possesses a 1-million-token context window, and is designed for coding, agent, and visual reasoning tasks.
Question: What hardware is required to run the model locally, and what is its approximate cost?
Answer: The model requires about 1 terabyte of memory, operating across two 512GB Mac Studio devices or 4 units of the 256GB version with the M5 Ultra chip, at an approximate cost of $38,000.
Question: Who contributed to bringing Meta and Apple together after a estrangement since 2017?
Answer: Daniel Gross, a former Apple executive and current Meta executive, played a pivotal role in reuniting the two companies following the 2021 privacy disputes and John Ternus taking the helm at Apple.
Question: How does this open model tie into Meta’s commercial strategy with the Muse Wonder model?
Answer: The open model trails the closed Muse Wonder model in performance, with Meta aiming through it to attract developers and enterprises to experiment locally with absolute privacy before scaling up to its larger services.
