01 / The short version
What happened
Meta released Muse Spark Open, an open-weight multimodal mixture-of-experts model with approximately 816 billion parameters and around 42 billion active per token. It supports text, images, audio, and video inputs with a context window of up to one million tokens. Meta describes it as coding-optimized for agentic workflows such as reading codebases, writing and debugging code, and tool use. Checkpoints include Muse-Spark-Open-816B-A42B and an NVFP4 variant for NVIDIA Blackwell GPUs. The BF16 checkpoint is about 1.5 TB in safetensors format.
See the exact references02 / Key takeaways
What you need to know
- 01
Muse Spark Open has approximately 816 billion parameters with around 42 billion active per token.
- 02
It accepts text, image, audio, and video inputs within a one-million-token context window.
- 03
Meta positions it as an agentic coding model trained on real developer workflows.
Keep in perspective
What to watch for
The source is a Meta announcement, so performance and capability claims are publisher-provided and not independently validated.
Go to the source
Exact references
These are the original pages used for this brief. Publisher claims are not independent evaluations.
01Primary source · Hugging FaceRead the original announcementhttps://huggingface.co/blog/meta-muse-spark-openAI-generated from the linked source. It can miss context; verify consequential details in the original. How the radar works