The development

What happened

Hugging Face’s 15 October integration post details Muse Spark Open, Meta’s large mixture-of-experts model for coding and extended agent workflows. It accepts several input modalities while generating text.

Key facts

  • The model has approximately 816 billion total parameters, with about 42 billion active per token and up to a one-million-token context.

  • Inputs include text, images, audio and video; multimodal input does not mean image or audio output.

  • BF16 and NVFP4 checkpoints are documented with Transformers, llama.cpp, SGLang and vLLM recipes.

  • BF16 weights occupy about 1.5TB. Activating only some experts per token does not remove the memory needed to store the others.

The context

Why it matters

TideReach assessment: One model can combine code with screenshots or recordings, but the full footprint makes deployment a substantial infrastructure decision. Smaller active compute is not equivalent to laptop-scale memory use. Open weights also do not by themselves establish unrestricted licensing.