Meta has released the weights for Muse Spark Open, a huge mixture-of-experts model built for coding, tool use and long-running agentic tasks. The company’s announcement lists 816 billion total parameters, 42 billion active parameters per token and a one-million-token context window.
The model card describes Muse Spark Open as natively multimodal: it accepts text, images, video and audio, then produces text. Meta says the model can handle multi-step reasoning, software-engineering work and extended tool-driven workflows. Those capability claims and the benchmark table come from Meta’s own evaluation.
The weights are available through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted access is also available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Supported inference software includes SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple’s MLX, llama.cpp and Unsloth.
A desktop model, if the desk holds two Mac Studios
Meta says Muse Spark Open is the first model in its class that can run on a desktop setup, though “desktop” is doing a lot of work there. Its reference configuration uses two M5 Ultra Mac Studios with up to 512GB of unified memory each, linked through MLX so the machines can hold the model together. Meta says it tested that setup on coding and video-editing workflows that ran entirely on-device.
