Meta launches Muse Spark Open, an 816B open-weight AI model for Macs requiring 1TB of RAM

Meta has launched Muse Spark Open, an 816-billion-parameter open-weight AI model designed to run locally on Apple Macs. It supports a context window of up to one million tokens, is natively multimodal, and targets complex coding and agentic workflows. Meta claims the model outperforms several popular open-weight alternatives in its evaluations.

The model runs through Apple's MLX framework, taking advantage of Apple silicon and unified memory. Meta demonstrated it on M5 Ultra Mac Studio hardware, which provides up to 1.2 TB/s of memory bandwidth per chip. However, running the full model reportedly requires at least 1 TB of memory and multiple Mac Studio systems, so local deployment isn't exactly cheap. For some perspective, a single 256 GB M5 Ultra Mac Studio costs $9,499.

Meta also tested the model with Muse Code, its AI coding terminal, demonstrating tasks such as fixing bugs, adding features, and building a Mastodon client. Other tests involved reviewing and cropping footage in DaVinci Resolve. As an open-weight model, Muse Spark Open supports customization and self-hosting, potentially reducing inference costs and keeping data on-device. It is also available through API platforms such as OpenRouter, Together AI, Fireworks AI, and Groq, with model weights available on Hugging Face.