New to this? Read it in simple words
- Meta released Muse Spark Open, a model that people can download and run themselves.
- It has 816 billion parameters, but only about 42 billion work on each token.
- An independent test gives it 39 points; Chinese open models still lead on coding.
- With Apple’s help, it runs on two or four linked Mac Studio computers.
- Open weights
- A trained AI model that people can download and run on their own computers, under the maker’s licence.
- Mixture of experts
- A model built from many smaller parts, of which only a few work on each piece of text.
- Token
- A small piece of text, often part of a word, that an AI model reads or writes.
Most of Meta’s recent Muse models could only be used through Meta’s own apps and services. On October 15, Meta released one that people can download and run themselves.
What Meta released
Meta calls Muse Spark Open “the most capable US open-weight model to date”. Open weights means that people can download the trained model and run it on their own computers. Downloads from Hugging Face need Meta’s approval first.
The model has 816 billion parameters, Meta says. It is a mixture-of-experts model, so only about 42 billion of them work on each token. It can read up to 1 million tokens at once. It takes in text, images, video and audio, and answers in text.
Companies can also use it through cloud services. Meta names Together AI, Fireworks AI and OpenRouter among others, and Amazon, Google and Microsoft offer the weights on their clouds.
Meta’s own figures of October 15; the score is from Artificial Analysis, the Mac setup cost is Engadget’s estimate.
Open, with conditions
The model is not open source. Its licence lets people use, change and share it, but products built with it must say “Built with Muse Spark Open”.
A company that sells access to the model through an API needs a separate agreement with Meta if its group earns more than $20 million in a year. The usage rules also ban military and weapons uses.
Sources3
How good it is
Artificial Analysis, an independent testing firm, gives the model 39 points on its Intelligence Index. The median for the models on its list is 19. It has no speed or price data for the model yet.
Meta’s own table shows a mixed picture. On Terminal-Bench 2.1, a coding test, Muse Spark Open scores 76.4. Two Chinese open models score higher in the same table: Kimi K3 with 85.0 and GLM 5.3 with 83.9.
“Coding is where it trails,” The New Stack wrote. Meta says the open version got extra safety training, and that its full safety evaluation report will be available soon.
Apple’s part
Meta says it worked with Apple so the model can run on Mac Studio desktop computers. A smaller, compressed version runs on two linked Mac Studios with 512 GB of memory each, or on four with 256 GB each.
The 512 GB Mac Studio goes on sale in late October, Engadget reports. It puts the cost of the four-computer setup at roughly $38,000.
Apple has not published an announcement of its own. Apple’s John Ternus praised the model on X. “Amazing to run a model of Muse Spark Open’s caliber and scale this efficiently, right from your desk,” Ternus wrote, as MacRumors reports.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1Primary source · October 15, 2026Introducing Muse Spark Open, The Most Capable US Open‑Weight Model to Date Meta
- 2
- 3
- 4Research · October 15, 2026Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta Hugging Face
- 5Research · October 15, 2026Meta ships Muse Spark Open, its answer to China’s best open-weight models The New Stack
- 6Research · October 15, 2026Meta built a huge open-weight AI model for Macs, but you’ll need 1TB of RAM to run it Engadget
- 7
- 8Research · Accessed October 15, 2026Muse Spark Open (Xhigh) - Intelligence, Performance & Price Analysis Artificial Analysis
The size, the licence and the benchmark table are Meta’s own. The only outside test so far is the Artificial Analysis score, and Meta’s full safety report is not published yet.
Meta’s release of a strong open-weight model opens real opportunities, though licence limits and trailing coding scores temper it. How we rate the mood