Apple-Meta Muse Spark Open: 816B AI on $10,799 Mac Studios

Apple and Meta are working together to let developers run Meta's new open-weight Muse Spark Open model locally on clusters of high-end Mac Studio desktops. Announced on Thursday, October 15, 2026, the collaboration pairs an 816-billion-parameter model with Apple's M5 Ultra hardware and MLX framework. It is a rare public partnership between two companies that have spent years clashing, and the catch is the price: according to CNBC, each Mac Studio in the required configuration costs at least $10,799, and you need more than one.

What happened

According to CNBC (Kif Leswing and Jonathan Vanian, published 1:00 p.m. Eastern on Thursday), Apple and Meta revealed a collaboration that lets developers run Muse Spark Open on Apple's high-end desktops instead of in a cloud data center. Customers can link between two and four Mac Studio computers and use the model for coding, video analysis, or other agentic work. The companies said running it locally adds a layer of security and saves users money. CNBC reports the announcement followed months of collaboration, and representatives for both companies declined to say whether it is the start of a deeper relationship.

Axios (Ina Fried) reported the same day that Meta is releasing Muse Spark Open as a model similar to Muse Spark 1.3, and that Meta calls it the most powerful U.S.-developed open-weight AI model available. As a proof of concept, Meta worked with Apple so the model runs across two or four connected Mac Studios powered by M5 Ultra chips, using Apple's MLX framework. Meta AI chief Alexandr Wang summed up the mood to Axios: "We're excited for mommy and daddy to be playing nice again."

Advertisement

Key details

The model: Muse Spark Open

  • Size and design: 816 billion parameters using a mixture-of-experts architecture, according to CNBC and Axios.
  • Meta's claims: Meta told CNBC it is the most capable open-weight model released under a license that allows limited commercial use. Meta told Axios it is the most powerful U.S.-developed open-weight model available. Both are company claims, not independent benchmark results.
  • License and access: Axios reports the model is free to download under a license that allows commercial use, and will also be offered through cloud providers and developer tools. CNBC describes the commercial allowance as limited, so teams should read the license terms before building a product on it.
  • Not Meta's top model: CNBC notes Muse Spark Open is less advanced than Meta's proprietary Muse Wonder, released earlier this week. Wang told CNBC these "aren't the most capable models that we've developed or released."

The Mac side: M5 Ultra and MLX

  • Ready for Apple silicon: CNBC says developers can download a version from sites like Hugging Face without extra configuration for Apple silicon, and that the model has been tuned to run faster on Apple hardware.
  • New Apple documentation: Apple revealed new chip benchmarks and published technical documentation on how its MLX framework can split open-source models across multiple machines, according to CNBC.
  • Clustering background: When Apple introduced the M5 Ultra Mac Studio in August, it said Thunderbolt 5 with RDMA lets several systems pool memory, and that a four-machine cluster delivers up to 3x faster AI inference than a single system in Apple's own testing.
  • Wang's pitch: "We worked with Apple to bring leading Agent capabilities to your desktop for the first time on the Apple M5 Ultra Mac Studios," Wang told CNBC.

The price tag

  • Per machine: CNBC reports a Mac Studio with the required M5 Ultra chip and 256GB of memory costs at least $10,799. Apple lists the M5 Ultra Mac Studio from $5,499, but that entry configuration is not the one CNBC says this model needs.
  • Multiple machines: Users need at least two. If each uses that 256GB configuration, a two-machine setup starts at about $21,600 before tax.
  • 512GB option: A more expensive Mac Studio with 512GB of memory arrives later this month, according to CNBC. Apple's September 22 availability note also lists the 512GB model for late October.

How the deal came together

  • Daniel Gross: Wang credited Gross, a former Apple director who now works at Meta, with playing a big role in bringing the two sides together, according to CNBC and Axios.
  • Macs inside Meta: CNBC reports Meta bought hundreds of Macs earlier this year so its Superintelligence team could experiment with local AI and agents.

Why it matters

  • Local AI gets a big model. Wang told CNBC that many powerful open-weight models are too big for most PCs, but Mac Studios offer enough memory to make Muse Spark Open practical. That gives developers a desktop route for a very large model.
  • Prompts stay on the machine. Because the model runs locally, Meta does not receive users' prompts, Axios reports. "Meta has no control or visibility to any of that," Wang said. That matters for companies wary of sending code or video to a third party.
  • A showcase for Apple silicon. Apple gets a flagship example of what its clustering and MLX tools can handle, which CNBC frames as a sign of Apple's growing strength in AI chips, an area where Nvidia has dominated so far.
  • Another Apple AI partner. Meta joins a list that already includes OpenAI, whose ChatGPT was integrated into Siri, and Google, whose Gemini models Apple uses in Apple Intelligence, CNBC notes. This deal is about running a third-party model on Apple hardware, not inside Apple's features.
  • A thaw between rivals. CNBC and Axios both describe a long history of friction between the companies, especially over privacy and ad tracking. Axios says the collaboration does not mean those differences are resolved, only that they no longer stop the two from working together when interests align.
  • The usual open-model caveat. Axios notes that making powerful models openly available raises concerns that bad actors could strip out safeguards.

What's next

Developers with M5 Ultra Mac Studios can download the model and try it across two to four machines. Independent reports on real-world speed and quality will matter more than either company's claims. The 512GB Mac Studio arriving in late October could change how many machines a setup needs. Wang told CNBC that Meta expects companies and developers to tinker with Muse Spark Open at small scale and then "ramp up and scale up their usage." Axios is watching a bigger question: whether Apple might eventually bring Meta models into Apple Intelligence. That is speculation, not an announced plan. For background on Meta's paid flagship, see our earlier coverage of Muse Wonder. We will update this story when independent benchmarks land or when Apple or Meta share more details.

FAQ

What is Meta Muse Spark Open?

Muse Spark Open is Meta's new open-weight AI model with 816 billion parameters and a mixture-of-experts design. Axios reports it is similar to Muse Spark 1.3 and free to download under a license that allows commercial use. CNBC describes that commercial use as limited.

Can Muse Spark Open run on a single Mac?

Not in the setup Apple and Meta announced. Both CNBC and Axios describe it running across two to four connected Mac Studios with M5 Ultra chips. CNBC says users will need multiple machines, and each 256GB M5 Ultra Mac Studio costs at least $10,799.

Does Meta see my prompts when I run it locally?

According to Axios, no. Because the model runs on your own hardware, Meta does not receive prompts or other information, and Wang said Meta has "no control or visibility" into how developers use it. If you use it through a cloud provider instead, that provider's own data terms apply.

Sources

Comments

Popular posts from this blog

OpenAI Agent Hacked an Australian Government Site: What We Know

Suncatcher TPU Satellite: How to Track It and What Happens Next

Gemini 4 Argon: How to Get Access (Plan, API Steps, Fairwind)