Meta Releases 816-Billion-Parameter AI Model Designed to Run on Mac Studios
Meta has launched Muse Spark Open, an advanced open-weight AI model designed for Macs, featuring 816 billion parameters and local execution capabilities. It excels in complex coding and agentic tasks, offering superior performance and the benefits of open-source transparency. The model is now available via various API platforms.
Meta has released Muse Spark Open, an open-weight artificial intelligence model designed for complex coding and multi-step tasks, with support for running locally on Apple’s high-end Mac computers. Announced on October 15, 2026, the model has 816 billion total parameters, 42 billion active parameters per token, and a context window of up to one million tokens, according to Meta.
Its natively multimodal architecture allows it to process text, images, audio, and video, while its agentic capabilities are intended to help developers complete extended tasks that involve planning, using tools, and correcting errors. Meta describes it as the most capable open-weight model developed in the United States to date, a claim based on the company’s own evaluations.
Running the model locally on a Mac requires substantial computing resources: Meta and Apple have demonstrated it on two Mac Studio computers equipped with M5 Ultra chips, providing approximately 1 terabyte of combined unified memory. The setup uses Apple’s MLX framework to support local inference, allowing developers to run the model on their own hardware rather than relying exclusively on cloud services.
Meta has demonstrated its capabilities through Muse Code, its AI coding terminal, including work on an open-source Mastodon client that involved fixing bugs, adding features, and building the application. The company also demonstrated the model operating DaVinci Resolve, where it reviewed video footage and selected and cropped clips.
The release highlights the growing appeal of open-weight AI models, which allow developers to download and deploy model weights and offer greater control over customisation and data handling than fully hosted proprietary systems. However, open weights do not automatically make a model inexpensive or easy to run: Muse Spark Open’s scale means local deployment requires considerable memory, compatible hardware and technical expertise.
Meta has made the model available through API providers including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, giving developers alternatives to operating it on their own infrastructure. Its practical value will depend on how well it performs across real-world workloads, alongside the cost and complexity of deployment.