Meta releases Muse Spark Open: a large AI model for programming that also runs on your own computers

On 15 October 2026, Meta released Muse Spark Open. It is a version of its AI model Muse Spark whose so-called weights are public. Those weights are the hundreds of billions of numbers that together determine how an AI model thinks and answers. Anyone who has them can download the model, run it on their own computers and train it further, without having to go through a Meta service.

Meta calls Muse Spark Open the most powerful American open model so far. The model is made for programming work and for agent tasks: jobs in which an AI independently makes a plan, uses tools, looks at the result and corrects itself. Besides text, it also understands images, sound and video. So it can look at a screenshot, listen to a voice message or follow a screen recording while it works on code.

Meta chooses open again

For years, Meta was one of the great champions of open AI models, with the Llama series. With the first Muse Spark models, from April 2026, the company chose a closed approach instead: they could only be used through Meta. In August, chief executive Mark Zuckerberg announced that Meta would release open models again. The small model Muse Glimmer, which can run on a laptop, was released immediately. Muse Spark would follow, and that has now happened.

Big, but clever with computing power

Muse Spark Open has about 816 billion parameters, the adjustable numbers that an AI model consists of. Yet for each piece of text it calculates with only a small part of them, about 42 billion. That is due to its design, which is called mixture-of-experts.

Compare it to a well-stocked workshop. There are 256 tools hanging on the wall, but for each job you only pick up 8. The workshop has to be big enough for all those tools, but you finish quickly because you only use what you need. It is the same with this model: it needs the memory of an enormous model, but works at the speed of a much smaller one.

The model can take in an enormous amount of information at once: about 1 million tokens (word pieces), enough for thousands of pages of text or a complete software project. Images, video and sound also fit into that working memory. The model always thinks for a moment before it answers. Users can set how long, from short for simple questions to very thorough for difficult programming jobs.

Running it yourself: from data centre to desk

The full version is about 1.5 terabytes in size and runs on heavy servers with powerful graphics chips. There is also a more compact version for Nvidia's newest chips. According to Meta, Muse Spark Open is the first model in its class that can run on a desk: on two linked Mac Studios with the M5 Ultra chip, each with 512 gigabytes of working memory. Meta had the model carry out real programming jobs this way on its own hardware, such as fixing bugs in an open-source app.

Hugging Face, the platform where the weights can be downloaded, tried out heavily compressed versions on a single Mac Studio with 512 gigabytes of memory. These reached about 20 tokens per second, faster than most people read. The team also showed that the whole model can be trained further in about two hours on 64 graphics chips.

Anyone without their own hardware can use the model through services such as Groq, Together AI and OpenRouter. The weights can be downloaded through Hugging Face, Amazon, Google Cloud and Microsoft.

How good is it?

Meta compares Muse Spark Open with other large open models, including the Chinese models Kimi K3, GLM 5.3 and Qwen. In that comparison, Muse Spark Open scores highest on a number of tests for agent tasks and on a very difficult knowledge test. On other tests, such as programming in the command line and a general intelligence index, Kimi K3 and GLM 5.3 score higher. The figures come from Meta itself. Meta will soon publish a full evaluation report, and independent testers will also examine the model themselves in the coming period.

Because an open model can be used without Meta's own safeguards, this version received extra safety training, according to Meta. In tests with attempts to misuse the model for cyberattacks, it refused more than 99 per cent of those requests, according to Meta. Meta advises companies to build in extra safeguards themselves when using it, such as a human check before an agent does something irreversible.

Free, with a few conditions

Meta is releasing the model under its own licence, the Muse Model License. Anyone may use, adapt, train further and commercially deploy the model for free. There are conditions, though. Products that use the model must state that they were built with Muse Spark Open. Companies that offer AI models as an online service and have annual revenue of more than 20 million dollars (about 18 million euros) must first sign a separate agreement with Meta. And the model is not intended for users under the age of 18. This means it is not open source in the strict sense, but it is a model that most developers and researchers can use freely.

What does this mean for the road to the singularity?

Muse Spark Open shows that a model that keeps up with the strongest open models in the world can now simply be downloaded and can even run on a desk. That is a good step forward. The more researchers and companies can work with such models themselves, the faster new applications, improvements and safety research will follow.

The model is also built for exactly the kind of work that makes AI increasingly independent: planning longer tasks, writing software and fixing errors. Such agents are an important building block if AI is ever to actively help develop better AI. The next step is for independent tests to show how strong the model is in practice, and what the open-source community will build with it. Given the pace of this year, we will not have to wait long for that.

What is the technological singularity? Read the explanation →