Meta has released Muse Spark Open, a very large AI model whose weights anyone can download. The company says it is “the most capable US open-weight model to date”, and that it can run on Macs.
The scale is real. So are the limits: the hardware is expensive, the licence has conditions, and the page that announces it gives no performance figures we could read.
What Meta released
Meta announced the model on 15 October. It is natively multimodal, meaning it handles more than text, and it is aimed at coding and “agentic” work: long tasks, complex software engineering and multi-step reasoning.
Meta’s page gives two sizes. The model has 816 billion parameters in total, and 42 billion of them are active. Our reading: in designs like this, only part of the model is used for each piece of text, which keeps the cost per answer lower than the total size suggests. It has a context window of 1 million tokens, which is roughly how much text it can keep in view at once.
Where to get it, and the licence
The weights are on Hugging Face, in the collection meta-models/muse-spark-open. Meta also lists Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. If you would rather not run it yourself, hosted versions are offered by Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
The licence is the Muse Model License. Meta says it has limited commercial terms and allows free use for many developers. “Open weights” is not the same as “open source”: you can download and run the model, but under Meta’s conditions. Read the licence before building a product on it.
Can you run it at home?
Technically, yes, but not on an ordinary computer. Meta’s own example is two Mac Studio machines with Apple’s M5 Ultra chip, linked and running the model through Apple’s MLX software. Each has up to 512 GB of unified memory, so together they come to about 1 TB.
TechCrunch, quoting Engadget, says the model needs at least 1 TB of RAM, with the best performance on a Mac Studio with an M5 Ultra. Meta’s page does not give a formal list of requirements. For other kit, if you are weighing a Mac purchase, see what to expect at Apple’s event on 28 October.
What Meta has not shown
Meta’s page has a section called Key Benchmarks, but we could not read any model scores in it. The only numbers are about speed: kernel speed-ups of 1.03 to 3.95 times on NVIDIA and AMD graphics chips, and a 3% gain end to end on AMD’s MI350.
The demonstrations are examples, with no metrics: fixing bugs and adding features in an open-source iOS app called Ice Cubes, and editing video in DaVinci Resolve. Any claim that it beats other open models is Meta’s own, and we found no independent test.
Safety
Meta says the model had extra safety training because open weights carry extra risk, and that it was evaluated under its Meta Superintelligence Scaling Framework and met the thresholds for release. The details are in a model card on Hugging Face, and Meta says the full evaluation report will come later.
For the earlier models in the same family, see our report on Muse Wonder, Meta’s first frontier AI models.
What we could not confirm
We have not run the model, and we found no independent benchmark. We could not open Engadget, the source of the 1 TB figure. We found nothing about UK availability or pricing beyond the downloads and hosted services listed above.
Leave a Reply