Meta gives away its big Muse Spark AI for free

Corrected 19 Oct: Muse Spark runs on two M5 Ultra Mac Studios (up to 512 GB each). All corrections

Meta has released the "weights" of a large AI model called Muse Spark Open. Weights are the saved numbers that make a model work, so anyone can download them and run the model. Meta says two Apple Mac Studio computers can run it together.

What happened

On October 15, Meta published a blog post saying it is releasing Muse Spark Open today. Meta calls it the most capable US open-weight model to date. It has 816 billion parameters in total, and 42 billion are active at a time. It can handle 1 million tokens of memory at once, which is roughly a very large pile of text. It works with more than text, and Meta says its focus is coding and tasks where the AI works through many steps.

TechCrunch reports that Meta says it beats other popular open-weight models, based on Meta's own tests. The article does not list the scores, and the Meta post also gave no numbers in the text we could read, so treat the claim as Meta's own.

On the Mac side, Meta says it is the first model in its class to run on a desktop, using two Apple M5 Ultra Mac Studios and Apple's MLX software.

Key facts

• Size: 816 billion parameters in total, 42 billion active, per Meta.
• Memory: 1 million tokens of context, according to Meta.
• Mac setup: Meta says each M5 Ultra Mac Studio can give its graphics chip up to 512 GB of shared memory, so two machines give about 1 TB.
• Download: Meta lists Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models.
• Hosted use: Meta lists Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.
• License: the weights come under the Muse Model License, which Meta says has limited commercial terms.

Background

An "open-weight" model is one whose files you can download. TechCrunch says these models can cost less to run and give more flexibility than closed models that live only on a company's servers.

TechCrunch also reports, citing Engadget, that the model needs at least 1 terabyte of RAM to run locally. That explains the two-computer setup. Meta says it tested the model with its coding tool, Muse Code, on jobs like fixing bugs in an iPhone app, adding features, and building an open-source Mastodon app called Ice Cubes. It also used it with video software to review footage and crop clips.

Meta says it added extra safety training because the weights are public, and that the model met its own release thresholds. The full evaluation report was marked "available soon."

What it means for you

Most people will not run this at home. It needs a lot of memory, and the TechCrunch article gives no prices for the Macs. You can try it more easily through the hosted services Meta lists.

Our take: the big news is access. A model this large used to sit behind a company's website. Now developers, small teams, and researchers can download it, change it, and run it on their own machines. That can mean lower costs and more privacy for people who have the hardware.

Our take: read the license before building a business on it. Meta says the terms are limited for commercial use.

What happens next

Meta says a full evaluation report is coming. Our take: watch for outside tests of the model, since the published scores so far come from Meta.

FAQ

Is Muse Spark Open really free?

Meta says many developers can use it for free under the Muse Model License, which has limited commercial terms.

What do I need to run it at home?

Meta says two Apple M5 Ultra Mac Studios with up to 512 GB of memory each. TechCrunch reports it needs at least 1 terabyte of RAM.

Can I use it without a Mac?

Yes. Meta lists hosted services like Groq, OpenRouter, and Together AI, and cloud platforms where the weights can be downloaded.

Sources

• research.meta.ai
• techcrunch.com