Meta releases open-weight model Muse Spark Open

Summary

Meta released Muse Spark Open, an open-weight mixture-of-experts model of roughly 816 billion total parameters with about 42 billion active per token, on Hugging Face. It accepts text, image, audio, and video inputs, supports a context window of up to one million tokens, and was trained for agentic coding such as reading codebases, writing and debugging code, and calling tools. Both BF16 (about 1.5TB) and NVFP4 checkpoints are downloadable; Meta's claim that it is the most capable US open-weights model to date is the company's own assessment.

Why it matters for our work

A frontier-scale agentic coding model can now be self-hosted, expanding options for organizations that cannot send code or data to external APIs. Multi-node GPU hardware is still required, so adoption decisions should account for operating costs.

Translated from the Korean original. Summaries may be translated and edited. Commentary reflects our perspective; forecasts remain the source’s views.