Apple Buried the Hatchet With Meta to Run Muse Spark Open

By TechPulse USA Newsroom — Cupertino, California | 15 October 2026

Update, Oct. 16

One day on, the independent picture is filling in. Per unite.ai’s launch writeup, the weights ship in two flavours — BF16 and NVFP4 — and running them locally needs a GPU with at least 700GB of VRAM for the NVFP4 version, or a full 2TB for BF16. And Meta’s own benchmark chart, broken down by ithinkdiff, is more mixed than the launch copy suggests: Muse Spark Open tops MCP-Atlas (89.2), JobBench (61.8) and HLE-Diamond without tools (26.1), but trails Kimi K3 and GLM 5.3 on the coding tests that matter most to its target users — Terminal-Bench 2.1 reads 76.4 against Kimi K3’s 85.0, and DeepSWE 1.1 scores 60 against Kimi K3’s 68. Those are Meta’s numbers; independent testing has yet to confirm any of them.

Meta and Apple just did the one thing nobody in this town thought they'd ever see: they worked together. On purpose. In public.

On Thursday, Meta released the weights for Muse Spark Open, an 816-billion-parameter open-weight AI model it calls the most capable US open-weight model to date — and revealed it had spent the last several months quietly collaborating with Apple to make the thing run, locally, on a Mac. Mark Zuckerberg announced it on X. Apple's CEO, John Ternus, replied with congratulations. If you've followed these two companies at all for the last decade, you know how surreal that sentence is.

And the collaboration isn't a press-release fiction. Two 512GB Mac Studio machines with M5 Ultra chips can hold the entire model in memory — what Meta described as effectively one supercomputer with a terabyte of unified memory, sitting on a desk. For an open-weights release, that's the headline inside the headline: frontier-scale AI you can run without a datacentre.

Photo: Meta CEO Mark Zuckerberg, who announced the Muse Spark Open release on X on Thursday. Photo by Anthony Quintano / Wikimedia Commons (CC BY 2.0).

What Meta actually put on the table

Muse Spark Open is a monster on paper: 816 billion total parameters, with 42 billion active per token — a mixture-of-experts design, meaning only a slice of the model fires for any given request. It handles a million tokens of context, is natively multimodal, and is built for coding and agentic work: the kind of model that doesn't just answer questions but goes off and does jobs.

The weights are downloadable from Hugging Face, with hosted inference through Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together AI, plus the big cloud channels: AWS SageMaker, Google Cloud's Model Garden, and Microsoft Foundry. In other words, Meta wants this model everywhere, immediately.

One housekeeping note, because the numbers are already getting mangled: Meta's materials say 42 billion active parameters per token. We checked this against a second outlet quoting 40B — Meta's figure is 42B, and that's the number we're printing.

This is Meta's second open-weights release since its 2025–26 AI overhaul, following Muse Glimmer in August — a 30B model released under Apache 2.0. Spark Open is a different animal, in scale and, as we'll get to, in licensing.

The real bombshell: Apple co-built this

Here's the part that would have been unthinkable three years ago. Meta says it worked with Apple "for several months" to get Muse Spark Open running locally on Mac Studio hardware with the M5 Ultra chip, using the latest version of Apple's MLX machine-learning framework. Apple updated MLX for the effort and published benchmark results on its ML Research blog — an unusually public engineering endorsement from Cupertino.

The demos Meta showed are genuinely wild. Muse Code — Meta's coding agent, powered by Muse Spark Open — made real iOS changes to Ice Cubes, the open-source Mastodon client: bug fixes, new features, a full build, shipped. In another demo, the model drove DaVinci Resolve end to end, reviewing video footage frame by frame and cropping it vertically. This isn't a chatbot demo; it's an agent doing actual desk work.

Photo: A Mac Studio (2022 model shown). Meta says two 512GB M5 Ultra Mac Studios can hold the entire 816B-parameter Muse Spark Open model. Photo by Yasu / Wikimedia Commons (CC BY-SA 3.0).

The hardware story matters because it's the whole point. Two 512GB M5 Ultra Mac Studios, each pushing 1.2 TB/s of memory bandwidth, hold the full 816B model. The 512GB Mac Studio ships later this month; the 256GB configuration starts at roughly $10,799, and a two-to-four-machine setup lands around $38,000. That is not pocket money. But it is a rounding error next to the cost of a GPU cluster — and it's a machine you can order, not a datacentre you have to build.

From courtroom enemies to… this?

Context, because the younger folks in the back may not remember: these two companies have spent years at each other's throats. Apple's App Tracking Transparency rollout in 2021 kneecapped Meta's ad-targeting machine — an estimated $10 billion hit. Meta has fumed about App Store fees and policies for just as long. Zuckerberg and Tim Cook have traded barbs in interviews for the better part of a decade. Meta is no stranger to privacy fights, either — the company recently denied that its Muse assistant reads private messages.

Which is what makes Thursday's cosiness so disorienting. Alexandr Wang — 29, ex-Scale AI, now Meta's chief AI officer, and apparently the industry's new unofficial couples therapist — told The Wall Street Journal in an interview: "This is an exciting moment for both companies, where we're finding ways to work together that are productive." He added: "We think that a lot of the world will run on open models, and they should be American."

Wang was even less guarded with Axios' Ina Fried, telling her: "We're excited for mommy and daddy to be playing nice again." He credited Daniel Gross — the former Apple executive now at Meta — with playing "a really big role in kind of getting the two companies to work nicely together."

Zuckerberg himself broke the news on X (@finkd): "Muse Spark Open — weights releasing today — runs great on new Mac Studios thanks to work with @johnternus and Apple." 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!" Apple, for its part, has been busy on the legal front too — a judge recently gave Apple access to OpenAI's internal files, and OpenAI has accused a former Apple hire of taking trade secrets.

Asked whether this signals a deeper relationship, both companies declined to say. Of course they did.

The licence, because there's always a licence

Now the part Meta would prefer you skim: Muse Spark Open is not released under Apache 2.0. It ships under a "Muse Model License" with limited commercial terms — a carve-out that means "open weights" here comes with strings attached. Read the licence before you build a business on it.

This is a genuine step back from Muse Glimmer's clean Apache 2.0 terms in August, and Meta deserves to be asked about it directly. "Open" is doing an awful lot of heavy lifting in this announcement.

Worth noting, too: Nvidia picked the same day to counter with its Nemotron 3.5 trio — Super VL at 120B parameters, the Bolt 74B-A9B running in 24GB at 2-bit precision, and the tiny Bolt E4B edge model at 4GB, due 23 October — under its OpenMDW-1.1 licence. The open-weights arms race just had its busiest Thursday in months.

Our take: two admissions wearing a partnership costume

Here's our honest read, because somebody should say it: this is the most interesting corporate détente in tech in years, and it's also two confessions dressed up as a celebration. Meta is admitting that winning the open-weights war requires distribution it doesn't own — actual computers, on actual desks, that developers actually buy. Apple is admitting that its AI story needs frontier models more than it needs the old feud. Both things can be true, and both are good for the rest of us.

The demos are the proof that matters. An agent that ships real iOS code and edits real video, running on hardware you can order this month, is a colossal step past the era of "here's a chatbot, good luck." The shoddy bit is the licence — limited commercial terms on a release marketed as open is the one sour note in an otherwise genuinely exciting day. Meta should fix that, or at least stop pretending the asterisk isn't there.

As The Wall Street Journal framed it, the release is also an answer to the lead Chinese open models have built — and on that front, an American model you can run on a Mac is a genuine counter-move. Whether this is a one-off engineering fling or the start of something real is the question both companies dodged. Watch what they do next, not what they posted today.

Muse Spark Open on Macs: your questions answered

What is Muse Spark Open?
Meta's new open-weight AI model: 816 billion total parameters (42B active per token), a million-token context window, natively multimodal, and built for coding and agentic tasks. Meta calls it the most capable US open-weight model to date. The weights were released on 15 October 2026 via Hugging Face and major inference providers.

How do I actually run it on a Mac?
You'll need serious hardware: Meta worked with Apple for several months to run the model locally on Mac Studio machines with the M5 Ultra chip, using Apple's MLX framework. Two 512GB M5 Ultra Mac Studios hold the entire model. The 512GB configuration ships later this month; the 256GB starts at roughly $10,799.

What does the licence actually allow?
Less than the word "open" suggests. Muse Spark Open ships under a "Muse Model License" with limited commercial terms — not Apache 2.0, unlike August's Muse Glimmer release. Read the licence carefully before building anything commercial on it.

Why would Meta and Apple team up now?
Mutual need. Meta gets its flagship open model onto premium developer hardware with Apple's engineering help; Apple gets a frontier AI story for the Mac Studio and a partner in the open-weights race. Both companies declined to say whether this signals a deeper relationship.

What's the catch?
Two: the price (a multi-machine setup runs around $38,000) and the licence (limited commercial terms). The model is open-weight, not open-source in the traditional sense — and that distinction matters.

Sources: Meta's announcement materials; The Wall Street Journal; Axios; MacStories; Apple's ML Research blog.

TechPulse USA is an independent tech news publication covering AI, chips, and Silicon Valley for a global audience. Our newsroom reports breaking tech stories with verified facts and original analysis.