Apple and Meta spent years as rivals. On Thursday, October 15, 2026, they announced a partnership anyway — and the reason is an AI model so large it needs two desktop computers working as one just to run. Meta released the weights for Muse Spark Open, calling it the most capable US open-weight model to date, and said it worked directly with Apple to tune the model for the Mac Studio with the M5 Ultra chip. For US developers, the headline is not just the size of the model. It is where it runs: on a desk, locally, instead of inside a distant cloud data center.
Key Takeaways
- Meta released Muse Spark Open on October 15, 2026 — an open-weight model with 816 billion total parameters, 42 billion active per token, and a 1-million-token context window.
- Apple and Meta worked together for months so the model runs locally across two to four Mac Studio machines with M5 Ultra chips, using Apple's MLX framework to split the work between machines.
- A qualifying Mac Studio is not cheap: a build with M5 Ultra and 256GB of memory starts at $10,799, and a 512GB version is expected later this month. Two linked machines can pool about a terabyte of unified memory.
- The model is natively multimodal — text, images, video and audio — and is aimed at coding and long, multi-step agent work, not casual chat.
- The weights are free to download from places such as Hugging Face, with API access through several providers, under a license that allows limited commercial use and adds extra conditions for large AI services.
What Happened: Rivals Announce a Joint AI Release
The announcement landed on Thursday through both companies. Mark Zuckerberg posted that the new weights were releasing that day and credited work with Apple CEO John Ternus and the Apple team. Ternus replied publicly, praising how efficiently a model of this scale can run from a desktop. Meta's AI chief Alexandr Wang said the two companies had collaborated for months, and pointed to Daniel Gross — a former Apple executive now at Meta — as one of the people who helped make the partnership happen.
That public warmth matters because this pairing has history. Apple and Meta had not publicly partnered since around 2017, when system-level Facebook integration disappeared from the iPhone. Their relationship then worsened after Apple's 2021 privacy changes limited ad targeting; Meta estimated in early 2022 that those changes would cost it about $10 billion in revenue that year. The leaders traded criticism for years after that. Neither company would say on Thursday whether this project is a one-off or the start of a deeper relationship — but shipping a flagship model together is a very different signal from the last decade.
The Model Itself: Big, Open and Built for Agents
Muse Spark Open uses a mixture-of-experts design. In plain English, the full model holds 816 billion parameters, but only about 42 billion are active for any single token. That keeps the cost of running it far lower than its headline size suggests, while still giving it a huge store of knowledge and skills to draw on. Its context window reaches 1 million tokens, enough room for very large codebases, long documents or extended project histories in a single session.
Meta is positioning it for work, not novelty. The company says it is built for coding, multi-step reasoning and extended agent tasks — jobs where an AI does a sequence of steps across software instead of answering one question. In its demonstrations, the model worked through Muse Code, Meta's AI coding terminal, on an open-source Mastodon client, handling bug fixes, new features and a full build. In another demo it drove the DaVinci Resolve video editor, reviewing footage frame by frame and preparing clips in a vertical format.
It also sits in a clear place inside Meta's own lineup. It is not Meta's most capable model — that role belongs to the proprietary Muse Wonder line, which we covered in our report on Meta's Muse Wonder frontier launch. Wang's pitch is different: a powerful open model that companies and developers can try at smaller scale on their own hardware, then scale up if it works for them. In August Meta released a smaller open model, and Zuckerberg had promised something bigger was coming. This is that bigger release.
Why Two Mac Studios Act Like One Supercomputer
The Apple side of the story is unified memory. Each M5 Ultra Mac Studio can let its integrated GPU use up to 512GB of memory, and the chip can move data at about 1.2 TB/s. Link two machines and, in Meta's description, they behave like a single desktop system with roughly a terabyte of unified memory — enough to hold the whole model. Apple's MLX framework handles splitting the model across machines, with distributed inference across two to four Studios. Apple also published technical documentation and benchmarks for that multi-machine setup alongside the launch.
The catch is price and timing. The 256GB M5 Ultra Mac Studio starts at $10,799 in the US, the 512GB version has not shipped yet and is expected later this month, and you need at least two machines for the full local setup. Meta is not pretending this is for every desk. Wang acknowledged that many powerful open-weight models are simply too large for most personal computers — Apple's high-memory Studios are the exception that makes this one practical. Meta itself bought hundreds of Macs earlier this year so its own teams could experiment with local AI and agents.
Why Local AI Is the Real Story for US Users
Running a frontier-scale model locally changes three things at once. First, privacy: prompts, code and files can stay on your own machines instead of being sent to a cloud service for every request. Second, cost control: after the hardware bill, there is no per-token cloud meter running while an agent works for hours. Third, independence: a local model keeps working on your schedule, under your rules, even if a provider changes prices, limits or access later.
That privacy point deserves a practical note for everyday users, too. Most people will never buy two Mac Studios, but the same principle applies at phone scale: the most private files are the ones stored locally, on your own device. That is why an app such as Pro Calculator Secure Vault keeps its secure vault for photos, videos, audio, documents and private notes in local on-device storage, with no data collected or shared with third parties. Local-first is becoming the theme of this AI cycle, from a $10,799 desktop to the Android phone in your pocket.
Who Can Actually Get Muse Spark Open, and on What Terms?
You do not need Apple hardware to try the model. Meta released the weights for download through Hugging Face and major cloud catalogs including Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. API access is offered through providers such as Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, according to launch coverage. That route needs no terabyte of memory — but it also gives up the full local-privacy benefit.
Open-weight does not mean no rules. The model ships under Meta's Muse license, which allows limited commercial use, requires clear attribution that a product is built with Muse Spark Open, and — as reported in launch coverage — requires AI services earning more than $20 million in a year from the model to notify Meta and sign a separate agreement before offering it as a service. Developers should read the license before building a business on it, rather than assuming open means unrestricted.
Before You Download: A Quick Reality Checklist
- Check the hardware math first: the full local setup needs two or more M5 Ultra Mac Studios and about a terabyte of pooled memory — a single ordinary Mac will not run it.
- Try the API route first if you are evaluating it for a team; it is the fastest way to test quality before committing to hardware.
- Read the Muse license carefully, especially the attribution rule and the extra agreement required for large commercial AI services.
- Download weights only from the official sources named at launch, such as Hugging Face or the major cloud catalogs.
- Watch for scams: fake "Muse Spark" download sites, repacked weights and paid "early access" offers are a predictable follow-on to a launch like this. Never pay an unofficial seller or run installers from social-media ads.
What This Means for Apple, Meta and the AI Race
For Meta, the release is proof its rebooted AI program can ship both proprietary and open models that other major companies want to work with — a confidence boost after a period when its efforts were seen as trailing OpenAI and Anthropic. For Apple, it is a hardware argument: if the most talked-about open model of the week runs best on Mac Studios, Apple's chips get credibility in a market long dominated by Nvidia's data-center hardware. It also fits Apple's broader pattern of partnering for AI — with OpenAI for ChatGPT in Siri and with Google for Gemini in Apple Intelligence — while its own newest Siri still faces criticism for lagging rival chatbots.
There is a policy angle, too. Open-weight AI has become a contested topic in Washington, where some officials worry free foreign models could undercut American AI. A high-profile US open-weight release, tuned for US-designed hardware, gives both companies a useful answer: openness, but on American terms.
Frequently Asked Questions
When was Muse Spark Open released, and by whom?
Meta released the weights on Thursday, October 15, 2026. Apple collaborated on optimizing the model for Mac Studio hardware, and both companies promoted the launch the same day, including public posts from Mark Zuckerberg and Apple CEO John Ternus.
How big is Muse Spark Open?
It has 816 billion total parameters with about 42 billion active per token, thanks to its mixture-of-experts design. It offers a 1-million-token context window and is natively multimodal across text, images, video and audio.
What computer do I need to run it locally?
Meta and Apple describe a setup of two to four Mac Studio machines with M5 Ultra chips. A 256GB M5 Ultra build starts at $10,799, a 512GB version is expected later this month, and two linked machines pool roughly a terabyte of unified memory. Anyone else can use the model through API providers instead.
Is Muse Spark Open free and open source?
The weights are free to download, but it is open-weight under Meta's own license rather than unrestricted open source. Limited commercial use is allowed, attribution is required, and large AI services above a reported $20 million revenue threshold need a separate agreement with Meta.
How is Muse Spark Open different from Muse Wonder?
Muse Wonder is Meta's proprietary frontier model, released earlier the same week. Muse Spark Open is the open-weight release: less capable than Meta's top proprietary model, according to Wang, but downloadable, inspectable and able to run on a customer's own hardware.
The Bottom Line
An Apple-Meta partnership would have sounded unlikely five years ago. On October 15 it produced the week's most interesting AI release: an 816-billion-parameter open model that turns two Mac Studios into a desktop supercomputer. The hardware bill keeps it out of most homes, and the license keeps it from being a free-for-all, but the direction is clear — powerful AI is moving from rented cloud time toward machines you own and data that stays with you. Expect more developers to test it through APIs first, more benchmarks in the coming days, and a close watch on whether this thaw between two old rivals produces anything bigger. You can find more apps built on that same local-first idea on the ZAS Technologies developer page on Google Play, including Pro Calculator Secure Vault with sign-in protection, an intruder selfie, an incognito browser and 8 smart calculators.
No comments:
Post a Comment