Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta
huggingface ·
Meta has released Muse Spark Open, an open-weight frontier multimodal model featuring a mixture-of-experts architecture with approximately 816 billion total parameters and 42 billion active parameters per token. Designed primarily for agentic coding tasks, the model supports text, image, audio, and video inputs within a context window of up to one million tokens. It utilizes a specialized perception encoder for visual data and a convolutional audio encoder to process diverse media types alongside text. The release includes BF16 and NVFP4 checkpoint formats, enabling deployment across various hardware configurations including NVIDIA Blackwell GPUs and Apple Silicon via quantization. Inference is supported through multiple frameworks such as Transformers, llama.cpp, SGLang, vLLM, and MLX, offering flexibility for different use cases. Additionally, the article details fine-tuning capabilities using the TRL library and demonstrates practical applications like 3D asset generation and browser automation agents. This release marks a significant step in open-weight AI by providing developers with a powerful tool for complex, multi-modal reasoning and execution.
huggingface