On October 15, Meta open-sourced the weights of the Muse Spark Open model — a multimodal AI model with 816 billion parameters, which the company calls the most powerful American open model to date. The highlight of the release is a partnership with Apple: the model can be run locally on two Mac Studio computers.
What Muse Spark Open is
Muse Spark Open is built on a mixture-of-experts architecture: of its 816 billion parameters, only 42 billion are active for each token. The model is multimodal from the ground up, has a context window of 1 million tokens and, according to Meta, was created primarily for programming and agentic tasks — long multi-step scenarios, complex software development and reasoning.
A trillion-parameter model on the desktop
Meta claims this is the first model of its class that can be run at home: this requires two Mac Studios with an M5 Ultra chip and 512 GB of unified memory each, or four machines with 256 GB each. The computers connect directly via Thunderbolt 5, and Apple's MLX framework distributes the model across them using RDMA technology. macOS 26.2 or newer and MLX 0.30.1+ are required, and a quantized NVFP4 version is used.
According to Apple Machine Learning Research, on a cluster of four Mac Studios the model generates more than 36 tokens per second — faster than a human can read. The Mac Studio configuration with 512 GB of memory is not yet on sale: reports say it is expected later this month.
Not just Mac
The model also runs on server hardware: on NVIDIA Blackwell accelerators (four GPUs for the NVFP4 version), AMD Instinct MI355X/MI350X and Intel Xeon 6 processors. The weights are available on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry, while the model is offered via API by Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. For fine-tuning, Meta recommends PyTorch TorchTitan.
Why this matters
As CNBC writes, the Meta–Apple collaboration lasted several months. For developers and companies, this means the ability to run a model on par with leading cloud services on their own hardware — without handing data over to third-party providers. At the same time, the release intensifies competition in the open-model market, where Chinese developers have so far dominated.
Sources: Meta AI Research, Apple Machine Learning Research, CNBC.