Meta Releases Muse Spark Open for Local Mac AI Workloads

Meta announced the release of Muse Spark Open on October 15, 2026. As part of the release, the company made the weights of the model publicly available for download. They describe the model as the "most capable open-weight AI model in the United States", which was created recently. It is tailored to solve coding tasks, complex software engineering tasks, multi-step reasoning, and agent activities.

What Makes Muse Spark Open Different

The released model allows developers to make the largest language model that is not based on cloud infrastructure. However, according to the company, for the full version of the model, it is necessary to have several high-performance computers instead of one.

Muse Spark Open has 816 billion parameters, 42 billion of which are used for the main calculations. The program simultaneously processes a million tokens and reads text, images, videos, audio, and PDF files. This model was created by Meta in cooperation with Apple, which provided its Mac Studio PCs with M5 Ultra chips to study and optimize the architecture of Muse Spark using the MLX machine learning framework.

According to the data provided, two Mac Studios were used as a test platform for the application. Each of them had 512 GB of unified memory, resulting in about 1 TB overall. Based on the results of the tests, the company claims that over 36 tokens per second were processed on four Mac Studios. It is important to note that this result is based on Apple’s internal test results and may vary for other Mac Studio configurations.

Where to Download Muse Spark Open and Its Licensing Terms

The weights of the model are available for download via the Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models platforms. It is noted that the Muse Model License has been approved for the weights, which allows some developers to use the model at no cost but with certain restrictions.

The company emphasizes that the model went through additional safety training due to the publicity of the weights. Meta mentions that the model met the requirements of the Meta Superintelligence Scaling Framework at the time of release; however, the company promised to publish a detailed evaluation report.

Performance, Hardware Requirements, and What to Watch Next

It remains unclear how good the model’s results are in practice, how many local computer resources are needed, and whether independent researchers will be able to replicate the results. It will be interesting to see the results of the evaluation reports and independent tests of the model’s performance on different hardware configurations.