Meta releases Muse Spark Open, an 816-billion-parameter open-weight model built to run on Mac Studios
Meta released the weights for Muse Spark Open on Thursday, calling it the most capable U.S. open-weight model to date. The coding- and agent-focused model can run locally on two Apple M5 Ultra Mac Studios after a rare collaboration with Apple, though the hardware costs at least $10,799 per machine.
What Meta released: Muse Spark Open is a natively multimodal model with 816 billion total parameters, 42 billion of them active at a time, and a 1-million-token context window, Meta said in its announcement. Meta calls it “the most capable US open-weight model to date” and says it is built for coding and agentic work such as long-horizon tasks and multi-step reasoning. The performance claims rest on evaluation results released by Meta itself, TechCrunch noted.
Running it at home: Meta says the model is the first in its class to run on a desktop, using two Apple M5 Ultra Mac Studios with Apple’s MLX framework. Customers can pair between two and four Mac Studios, CNBC reported. A Mac Studio with the required M5 Ultra chip and 256GB of memory costs at least $10,799, and a 512GB version is due later this month, CNBC said. TechCrunch, citing Engadget, said running the model locally requires at least 1 terabyte of RAM.
Where developers can get it: The weights are available for download on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, and the model is offered through API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, Meta said. It is released under the Muse Model License with limited commercial terms. Meta said the model received additional safety post-training because it is open-weight and met the thresholds of its Meta Superintelligence Scaling Framework for release, with a full evaluation report “available soon.”
Apple and Meta: The release is a rare public collaboration between two longtime rivals, which had not struck any public partnerships since 2017, CNBC reported. Meta AI chief Alexandr Wang told CNBC that Muse Spark Open is not as advanced as Meta’s proprietary Muse Wonder model. “These aren’t the most capable models that we’ve developed or released,” Wang said.
Why it matters now: Open-weight models let developers download a model’s parameters and run it on their own hardware, which can lower costs and keep work off the cloud, CNBC and TechCrunch noted. Meta’s own demos, including fixing and building an open-source Mastodon app and driving DaVinci Resolve video editing, show the agent pitch, but the hardware bill means running a model this size at home remains an expensive proposition.