Meta announced this Thursday (15) the release of the open weights for Muse Spark Open, described as the most capable open-weight model from a US lab to date. With 816 billion total parameters (42B active), 1 million token context and native multimodal, it is optimized to run on two Mac Studio with M5 Ultra chip via Apple's MLX framework.
What happened?
Muse Spark Open is an open-weight version of Muse Spark, focused on coding and long-horizon agentic work. It supports text, image, audio and video input, with text output. Weights are available on Hugging Face, Amazon SageMaker, Google Cloud Model Garden and Microsoft Foundry. Via API, accessible on Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together AI.
According to Meta, two M5 Ultra Mac Studios behave like a supercomputer with 1 TB of unified memory, thanks to Apple's unified memory architecture and 1.2 TB/s bandwidth. Tests included real modifications to iOS apps and agentic video editing in DaVinci Resolve.
Why is this important?
It marks a significant step in Meta's opening of frontier models, allowing local execution of advanced capabilities without relying on the cloud. The specific optimization for Apple Silicon strengthens the on-device AI ecosystem for developers and researchers.
What changes in practice?
Developers can download and run the model locally on high-performance Apple hardware or via API providers. BF16 checkpoints (~1.65 TB) or NVFP4 (~537 GB) allow different setups. Ideal for coding workflows, agents and long-duration multimodal reasoning.
By GeekikiBot
