Meta Muse Spark Open Is Out: A Huge AI Model on Your Desktop

Public institutions, law firms and large companies evaluating an AI tool usually start by asking where their documents end up. Muse Spark Open, which Meta released on October 15, is worth a look from that angle: Meta says it is the first model in its class that can run on a desktop setup of two Mac Studios. That setup needs about a terabyte of memory and tens of thousands of dollars in hardware, and the license comes with conditions. This post looks at who the model might make sense for and where its limits are.

What happened

Meta Superintelligence Labs released the weights for Muse Spark Open and introduced it as the most capable US open-weight model to date. Open weights means you can download the model and run it on your own hardware.

The model uses a mixture-of-experts architecture with 816 billion total and 42 billion active parameters, and a 1-million-token context window. It takes text, image, video and audio as input and produces text. Meta positions it for coding and agentic work. The weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry, and the model is also available through API providers such as Together AI, Fireworks, Groq and OpenRouter.

The work with Apple is about how the model runs locally. It was optimized for Apple’s MLX framework. According to Meta, the GPU in each M5 Ultra Mac Studio can use up to 512 GB of memory, so two machines together offer about a terabyte of unified memory and can hold the full model.

The Verge describes the two companies as typically rivals. Zuckerberg thanked Apple’s new CEO John Ternus on X, and Ternus replied with congratulations. Meta’s chief AI officer Alexandr Wang told Axios, per The Verge, that they’re “excited for mommy and daddy to be playing nice again.” CNBC notes the two haven’t announced a public partnership since 2017. The press has read this as two rivals warming up, but for now it is limited to this model; according to CNBC, company representatives declined to say whether it marks the start of a deeper relationship.

What it means for your brand

According to CNBC, the companies say running the model locally rather than in the cloud adds a layer of security and saves users money. The security side is straightforward. When the model runs on your own machines, contract archives, tender files, internal correspondence or visuals for an unreleased campaign can be processed without being sent to a cloud provider. The savings depend on usage volume, because, as the next section shows, the hardware is expensive.

That makes it an option worth evaluating for organizations that struggle to send documents outside. In the public sector, personal data and official documents going to third-party servers can cause problems in procurement and legal reviews. For law firms, client confidentiality is one of the main reservations about cloud tools, and a locally hosted model could be tried for internal work like document summaries and research drafts. Enterprise brands could analyze pre-launch product information, archive footage and internal reports without sending them outside. Two questions stay open for all of them. How well does the model handle documents in their language? Meta hasn’t published data on Turkish, for example. And does the setup cost pay off at the organization’s usage volume?

One of the workflows Meta tested involves video. The model drove DaVinci Resolve end to end, reviewing footage frame by frame, picking the strongest moments and cropping them vertically. Organizations with large, sensitive video archives could try a similar flow for a first-pass selection without sending unreleased footage anywhere. This was Meta’s own test, and only testing on your own archive will show how useful it is. Building a story from the selected moments and setting the rhythm and the brand’s voice is still the job of a video production team.

Costs and constraints

Hardware

Engadget points out that running the model locally takes roughly 1 TB of memory. The two-machine setup Meta describes relies on the 512 GB Mac Studio. CNBC reports that this version arrives later this month; per Engadget, its price hasn’t been announced. The 256 GB M5 Ultra Mac Studio on sale today starts at $10,799, according to CNBC. Reaching the same memory with that version takes four machines, and Engadget estimates such a setup at roughly $38,000. Outside the Mac the picture is similar: per the model card, you need at least 700 GB of GPU memory for the NVFP4 version and 2 TB for BF16.

License

Meta describes the Muse Model License as having limited commercial terms. Per the license text:

• Every product or service using the model must prominently display “Built with Muse Spark Open” in the interface, documentation, website or settings menu.
• If an organization with more than $20 million in annual revenue offers the model to others as a service via API, it needs a separate agreement with Meta. Internal use is exempt from this clause.
• The usage policy prohibits military use and espionage, operating critical infrastructure, and the unauthorized or unlicensed practice of professions such as law, finance and medicine.

That last point bears directly on the public-sector and legal uses above. Have your legal team read the text before deciding on a setup.

Performance and language

Muse Spark Open isn’t Meta’s strongest model. Speaking to CNBC, Wang acknowledged these aren’t the most capable models Meta has built, and CNBC notes that its proprietary Muse Wonder is further ahead. Language support is uncertain too. Meta says the model was trained on data from more than 100 languages but hasn’t been evaluated on all of them, and there’s no published data on Turkish.

Practical steps

1. Decide which documents must never leave your premises and which can go to a cloud API. Your need for a local model follows from that list.
2. Compare your monthly cloud token spend against the hardware investment and upkeep. Engadget writes that for a large organization burning through its token budget, a few Mac Studios might look like a good deal, provided the model performs as well as Meta says.
3. Before buying hardware, test the model through an API provider on non-sensitive sample documents, in your working languages. If the results are good enough, then talk hardware.
4. Have your lawyer read the license. The attribution requirement, usage policy and model-as-a-service clause can play out differently for each organization.
5. Add human approval to what the agent does. The model card recommends extra safeguards such as human-in-the-loop confirmation for irreversible actions. A draft prepared by the agent, for example, shouldn’t go out until a responsible person approves it.
6. If you connect the model to a client portal or website, plan where the required attribution will appear from the start. That decision belongs in the product design process.

We covered Meta’s Muse shopping-agent announcements in our Meta Muse post. This post is about running the model on your own infrastructure.

FAQ

Can we use Muse Spark Open for free in a commercial project?

According to Meta, the license allows free use in many cases. The conditions are the “Built with Muse Spark Open” attribution, compliance with the usage policy, and a separate agreement with Meta for organizations above $20 million in annual revenue that offer the model as a service. Leave the final call to your legal team.

Does it run on a single computer?

No. The desktop setup Meta describes uses two M5 Ultra Mac Studios with 512 GB of memory each, about 1 TB in total. CNBC says two to four Mac Studios can be paired; with the 256 GB version on sale today, you need four.

Does our data really never leave the premises?

When the model runs on your own machines, your prompts don’t go to a cloud provider. Security doesn’t stop there; who can access the machines, which tools the agent is connected to and how the network is configured also matter. And to download the weights from Hugging Face, you have to submit personal and organizational details through an access form.

If you’d like to work out where AI could help your organization, which data needs to stay in-house, and how that affects your website, your video archive and your internal tools, get in touch.