Meta released Muse Spark Open on Thursday, the most capable model it has ever let anyone download, and says it runs on a pair of Apple Mac Studios. Two numbers the announcement leaves out decide what that means for a small office. Renting the model from any of the four companies hosting it on OpenRouter costs $1.25 per million tokens in and $4.25 per million out, exactly what Meta charges for its closed Muse Spark. And the Mac Studio that Meta’s two-machine setup depends on, the one with 512GB of memory, is listed on Apple’s store today as “coming late October,” with no price.
The short version: for a five-person office doing a steady day of document work, renting Muse Spark Open costs about $141 a month by the arithmetic below. Owning it means at least $10,998 in hardware before Apple has even priced the version that can hold it, a floor that takes six and a half years of that rent to pay back. The good news is real anyway. Now an office that cannot let certain files leave the building no longer has to settle for a small model to keep them there.
What is Muse Spark Open?
Meta Superintelligence Labs published Muse Spark Open on October 15 and calls it “the most capable US open-weight model to date.” It has 816 billion parameters, of which 42 billion do the work on any given token, and it reads a context of up to one million tokens. It takes in text, images, video and audio, and it was built for coding and for agent work: long, multi-step jobs where the model uses tools and checks its own progress. The weights are on Hugging Face and in the Amazon, Google and Microsoft cloud catalogs.
Neither of Meta’s two demonstrations is a chat: in one its Muse Code tool changed an open-source iPhone app, in the other the model drove the video editor DaVinci Resolve to find and crop vertical clips. On Meta’s own benchmark table it scores 80.4 on OSWorld Verified, a test of operating a real computer, and 76.4 on Terminal-Bench 2.1. Those are the vendor’s numbers, and the strongest model Meta has, Muse Wonder, stays closed, as Bloomberg reported.
What does Muse Spark Open cost to rent?
OpenRouter, which lists prices across hosting companies, showed four hosts for the model when we read it on October 15: Together, Nebius, Fireworks and Baseten. All four charge $1.25 per million input tokens and $4.25 per million output tokens, and three of them charge $0.15 per million for input they have cached. That is identical, to the cent, to Meta’s own closed Muse Spark 1.3 on the same page. Free weights have not made the hosted version cheaper, at least not on day one.
To turn that into a monthly figure, take a five-person office where each person runs 40 AI tasks a workday, each task sending 20,000 tokens of documents and instructions and getting 2,000 tokens back. That is 4 million tokens in and 400,000 out per day: $5.00 plus $1.70, or $6.70 a day. Over 21 workdays it comes to about $141 a month, before any caching discount. Your own numbers will differ; the method will not.
One thing that does not carry over from the closed model is the cheap lane. Meta’s Contributor tier cuts Muse Spark to $0.10 and $0.20 per million tokens if you let Meta train on your content. No such tier appears for the open model on OpenRouter, where independent hosts run it rather than Meta.
What would it take to run Muse Spark Open in your office?
Meta’s recipe is two Mac Studios with the M5 Ultra chip and 512GB of memory each, linked together and running Apple’s MLX software. The model card explains why one will not do: the compressed version needs at least 700GB of memory. You can walk out of an Apple Store with a Mac Studio today. You cannot walk out with the one this model needs, and you need two.
Here is what Apple’s store showed on October 15. The M5 Ultra Mac Studio starts at $5,499 with 96GB of memory. The largest memory option you can order today is 256GB. The 512GB option carries the note “coming late October” and no price. So the honest floor is two base machines at $10,998, a figure that cannot actually run the model and will only go up once the 512GB price appears.
Speed has a ceiling you can work out from Apple’s spec sheet. Every token the model writes, roughly a short word, requires reading its 42 billion active parameters from memory, about 24GB at the compressed size. The M5 Ultra moves 1.2TB of memory a second, so with the model split across two machines taking turns, the best case is about 50 tokens a second. Real speed lands below that, and that pair of machines serves the whole office, not one desk.
Set side by side, the three paths look like this.
| Path | Cost, read October 15 | The catch |
|---|---|---|
| Rent from a host | $1.25 in and $4.25 out per million tokens; about $141 a month for the five-person example | Your documents travel to the host |
| Own two Mac Studios | At least $10,998, and more once the 512GB option is priced | The configuration Meta names is not on sale yet |
| Wait | $0 | None; the hardware price arrives in about two weeks |
At $141 a month, the $10,998 floor takes 78 months to pay back, before electricity and before the real 512GB price. Nobody should buy these machines to save money. The only reason to buy them is the one that has always justified local AI.
Who should own a model like this?
In August we wrote that the whole case for running AI on your own hardware is the unwritten list of work you refuse to upload: client financials, patient intake, signed contracts. That case has not changed. What changed on Thursday is the quality of the model on the other end of it. The open model Meta offered then, Muse Glimmer, was a 30-billion-parameter system built to fit on one graphics card. Muse Spark Open is 27 times larger and built for long agent jobs. A small accounting practice or a two-provider clinic with real confidentiality obligations used to choose between privacy and capability. It can now have both, for the price of two computers.
That is also a turn from a few months ago, when the largest open model of the moment was one you would never actually download because it needed a datacenter. This one needs a shelf.
What does the Muse Spark Open license require?
Muse Glimmer shipped under Apache 2.0, one of the most permissive common licenses. Muse Spark Open comes under Meta’s own Muse Model License, which is still generous: commercial use, fine-tuning and resale are allowed, and Meta claims no ownership of what the model writes. Two conditions are new. Any product or service built on it must “prominently display” the line “Built with Muse Spark Open” in its interface, so a customer-facing chat on your website would carry that label. And a company with more than $20 million in revenue that wants to resell the model as a hosted service needs a separate agreement with Meta, a threshold most readers here will never reach.
Should a small business do anything this week?
Three sensible moves, depending on where you sit:
- If you have that list of untouchable work, write down your real monthly volume now, then price the 512GB machines when Apple publishes the figure in late October and run the payback test with your own numbers.
- If someone builds AI tools for you, ask whether Muse Spark Open is on their list. It costs the same as Meta’s closed model and four companies already host it, so the tool is not tied to a single provider.
- If neither applies, keep using what you rent. This release is good news you can leave alone.
Frequently Asked Questions
What is Muse Spark Open?
Muse Spark Open is an open-weight AI model Meta released on October 15, 2026, with 816 billion parameters, 42 billion of them active per token, and a one-million-token context. It handles text, images, video and audio and was built for coding and agent work. Anyone can download the weights from Hugging Face under Meta’s Muse Model License.
How much does Muse Spark Open cost to use?
The weights are free to download. Hosted access on OpenRouter cost $1.25 per million input tokens and $4.25 per million output tokens across all four listed hosts on October 15, 2026, the same as Meta’s closed Muse Spark 1.3. A five-person office running 40 document tasks per person each workday would pay about $141 a month.
Can Muse Spark Open run on a Mac?
Meta says it runs on two Mac Studios with the M5 Ultra chip and 512GB of memory each, linked with Apple’s MLX software. On October 15, 2026, Apple’s store listed the 512GB option as coming in late October with no price, and the base M5 Ultra model with 96GB started at $5,499, so two machines cost at least $10,998.
Can a small business use Muse Spark Open commercially?
Yes. The Muse Model License allows commercial use, fine-tuning and resale, and Meta claims no ownership of outputs. Products built on it must prominently display “Built with Muse Spark Open” in their interface, and companies with more than $20 million in revenue that offer it as a hosted service need a separate agreement with Meta.
If you could keep a model this capable on a shelf in your own office, which job would you hand it first?
