Muse Spark Open: top-tier AI to download, 700 gigabytes heavy

TL;DR. On October 15, Meta released Muse Spark Open: an AI model with 816 billion parameters, built for coding and agents, that any company may download and run itself. Meta calls it the strongest open model from the US. On the independent Artificial Analysis index it scores 39 points, the best open model scores 46 and the best closed one 58. And it needs at least 700 gigabytes of GPU memory. It is too big for a mid-sized company's server room. But it shows that you now have a real choice when you ask: who runs our model?

What Meta released

Open weights means Meta publishes the trained numbers of the model for download. Whoever has them can run the model on their own hardware, adapt it, or have it run by a provider of their choice. The key facts according to Meta's announcement and the model card:

• Large, but frugal in computing. Of the 816 billion parameters, only 42 billion work on each word. The model consists of 256 specialist parts, of which 8 are asked each time.
• Lots of context. Up to one million tokens, several thick binders at once. It reads text, images, video and audio and answers in text.
• Built for agents. It plans steps, calls tools, checks the result and corrects itself. On the tool test MCP-Atlas it scores 89.2 points according to Meta, ahead of every open model in the comparison.
• Available everywhere. Downloads on Hugging Face and in the clouds of Amazon, Google and Microsoft. As a ready-made service, Together AI for example charges $1.25 per million input tokens and $4.25 per million output tokens.

Meta keeps its stronger models Muse Wonder and Muse Wonder Full closed. How far the open model reaches is shown by the Artificial Analysis index, which combines many individual tests into one number:

The catch: 700 gigabytes

Nothing about this model is small. Even in its compressed version it needs at least 700 gigabytes of GPU memory according to the model card, and 2 terabytes at full precision. Meta demonstrates the model on two linked Mac Studios with 512 gigabytes of memory each. Apple says that configuration will only arrive at the end of October, and there is no price yet. Meta has not published measured speeds on the two Macs, VentureBeat reports.

For comparison: Nvidia's Nemotron 3.5 Bolt, also available since October 15, fits into 24 gigabytes, about a thirtieth of the memory. Nvidia even publishes the training data. According to Nvidia, it compares with models of similar size, not with the top tier.

The fine print

• Open does not mean without conditions. The Muse Model License requires every product using the model to display “Built with Muse Spark Open” prominently: in the interface, the documentation, on the website or in the settings.
• The 20 million threshold affects very few. Only those who sell the model to others as an AI service and make more than $20 million in revenue over twelve months need a separate agreement with Meta. Internal use and building it into your own product are exempt.
• The download comes with a form. Meta asks for your full legal name, date of birth and company details.
• The numbers come almost entirely from Meta. Meta has only announced its detailed evaluation report. On the coding test Terminal-Bench, the Chinese model Kimi K3 is ahead, 85.0 to 76.4 points.
• Not every provider delivers the same. Groq offers 131,072 tokens of context, Fireworks the full million.
• Safety is up to whoever runs it. Meta gave the open version additional training against misuse, but warns: modifying the model can weaken these safeguards. According to the model card, irreversible actions should be confirmed by a human.

What this means for a mid-sized company

Until now the question was usually answered quickly: if you wanted the best model, you sent your data to one of the big providers. Open models of this class change that. You decide who runs the model: your own server, a data center of your choice, or a service that offers it ready to use. For contracts, design data or personnel files that should not leave your company, that is a real difference.

The price is the distance to the top. On the index, the best open model is 12 points behind the best closed one, Muse Spark Open is 19 behind. For the hardest tasks, such as long programming jobs across many files, closed models stay ahead. For many tasks in day-to-day operations you do not need the very top. Which ones those are for you, only a test with your own tasks will show. And the bill changes: if you run it yourself, you no longer pay per word, but for hardware, power and someone who looks after it.

What you should do now

1. Sort your data. Decide which data may go into someone else's cloud and which may not. Only for the second group is it worth asking about running a model yourself.
2. Rent first, then buy. Test an open model through a hosted service first, with 30 to 50 real but non-sensitive tasks from your daily work. Only when the results hold up is it worth doing the math on your own hardware.
3. Start with a small model. For sorting, searching and extracting, a model that fits on a single computer is often enough. You only need the big one for long, multi-step tasks.
4. Have the license reviewed. The attribution requirement, prohibited uses and passing it on to customers should be settled before your first product, not after.
5. Put a human in front of the last key. Whatever an agent cannot undo is confirmed by a person. Meta recommends this itself.

Muse Spark Open is not a model for every basement. But it shifts the question from “Which provider?” to “Where should our data live?”. If you would like to think that question through calmly for your company, we are happy to talk.