Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta

Meta has introduced Muse Spark Open, an open-weight frontier multimodal mixture-of-experts model featuring approximately 816 billion parameters. According to Hugging Face, the model activates around 42 billion parameters per token and supports a context window of up to one million tokens. Following the smaller Muse Glimmer model released in August, Muse Spark Open scales up the mixture-of-experts architecture to handle complex developer workflows, including agentic coding, reading codebases, debugging, tool calling, and browsing the web with images. The model natively processes text, images, audio, and video inputs, allowing users to interact via screenshots, voice notes, or screen recordings alongside code tasks. The architecture features separate input embedding and output layers of about 1.66 billion parameters each, 62 language model layers with grouped-query attention, and a Mixture-of-Experts block containing 256 experts where a router selects 8 experts per token. For perception, an encoder handles images and video frames, while an audio encoder processes raw waveforms using a convolutional pyramid. Deployment options include BF16 and NVFP4 checkpoints, the latter optimized for NVIDIA Blackwell GPUs to run on a single B200 node. The processor suite includes tools for handling chat templates, image patch expansion, video frame and audio downsampling, and adjustable reasoning effort levels.