Meta shares open weights for its massive Muse Spark coding model

Meta has released the open weights for Muse Spark Open, an 816-billion-parameter multimodal artificial intelligence model that had been proprietary since April. Built primarily for coding and long-running autonomous agent tasks, the system features a context window spanning up to one million tokens and can process text, image, audio, and video inputs to generate text. [1][3]

The model uses a mixture-of-experts architecture with 256 experts per layer, activating just 42 billion parameters per token to accelerate inference speeds while demanding substantial memory to keep all weights loaded. Meta released checkpoints in standard BF16 format—which weighs roughly 1.5 terabytes—and an NVFP4 four-bit precision format designed for Nvidia Blackwell GPUs. Through a partnership with Apple, the model was also fine-tuned for consumer hardware, managing a throughput of 20 tokens per second across two Apple M5 Ultra Mac Studios. [2][3]

Meta bills the release as the most capable open-weight model produced by an American laboratory to date. By opening up weights optimized for complex tasks like multi-file refactoring, web browsing, and tool execution, the company provides developers with an open alternative to proprietary frontier systems. [1][3]