Meta Releases Muse Spark Open: An 816B Parameter Multimodal Model
Meta has released Muse Spark Open, an 816B-parameter native multimodal open-weight model featuring a 1M-token context window designed specifically for code generation and agent collaboration. The model weights are now publicly available on Hugging Face and other major platforms under a restrictive commercial license. Claimed by Meta to be the most capable US open-weight model to date, it sets a new benchmark for open-source AI in handling long-range tasks and complex software engineering. Its ecosystem-wide support and local execution capabilities democratize advanced AI research and enterprise deployment. Muse Spark Open utilizes an 816B total parameter architecture with 42B active parameters, and can be run locally on desktop environments using two Apple M5 Ultra chips thanks to the MLX framework. It is also compatible with mainstream inference engines like vLLM, SGLang, and TensorRT-LLM.
## BACKGROUND
Open-weight AI models provide developers with model weights to run, fine-tune, and adapt locally, contrasting with closed API-only models. A context window measures the amount of text a model can process at once, while frameworks like Apple MLX optimize machine learning workloads specifically for Apple silicon unified memory.