Meta releases massive AI model that doesn’t need data center

Meta on Thursday released the weights for Muse Spark Open, an 816-billion-parameter artificial intelligence model that the company says can run on two Apple Mac Studio computers instead of in a data center.

The company called Muse Spark Open the “most capable US open-weight model to date.” It is built for coding and agentic work, meaning AI that carries out multistep tasks on its own. An open-weight model can be downloaded and run on a user’s own hardware rather than accessed only through a company’s servers.

Muse Spark Open uses a mixture-of-experts design, so only 42 billion of its parameters are active for each token it processes, which reduces the computing power needed. The model is natively multimodal and has a context window of 1 million tokens, according to Meta.

Meta said Apple’s M5 Ultra chip allows each Mac Studio to use up to 512 gigabytes of unified memory. Two machines can run a compressed, 4-bit version of the model using Apple’s MLX machine-learning framework, according to Apple’s research. In its native precision, the model would need more than 1.6 terabytes of memory.

The hardware is expensive. Macs with at least 1 terabyte of memory combined would cost more than $40,000, The Wall Street Journal reported, as relayed by Quartz. Apple’s 512-gigabyte configuration is not yet on sale and is due later this month, CNBC reported.

Developers can download the model in two formats from Hugging Face.

The release enters a crowded open-weight field. Nvidia published Nemotron 3 Ultra, which has 550 billion total parameters and 55 billion active, in June, along with its training data. Chinese developer DeepSeek’s V4-Pro, released in April, has 1.6 trillion total parameters and 49 billion active.

The release also marks a shift for Meta. When it launched Muse Spark in April, the model was closed, a break from the open Llama releases that preceded it.