Meta has introduced Muse Spark Open, the most powerful open-weight model from the US. It is a model with 816 billion parameters, with a context window that can use one million tokens at a time. The model is primarily built for programmers and orchestration of AI agents, which then work independently. Interestingly, Meta has combined this model with Apple optimized for the MLX framework and the latest computers Mac Studio and chip M5 UltraIf you have a couple Mac Studio with 512 shared memory, i.e. 1TB of memory in total, then you can run the model locally, without having to connect to Meta servers. Apple during tests, he stated that if he used four of the latest Mac Studio with a chip M5 Ultra together, the model exceeded a speed of 36 tokens per second.
Muse Spark Open is available today through API platforms such as Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Model weights can be downloaded from the platforms Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models. The model works on leading inference engines such as SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, including optimizations for AMD hardware, Apple, Intel and NVIDIA.
