Nvidia offers Muse Spark Open weights and NIM deployment container

Nvidia says developers can download Meta's Muse Spark Open in NVFP4 format from Hugging Face and deploy it through a downloadable Nvidia NIM container. The container automatically selects a serving backend and numerical precision for the detected hardware. The model has 816B total parameters, activates 42B per token and supports a 1M-token context window. Nvidia says the NVFP4 checkpoints reduce its memory footprint on Blackwell Ultra hardware. Developers can also fine-tune the model through NeMo AutoModel using Hugging Face checkpoints without conversion. Nvidia claims higher throughput and lower cost per token, but supplies no numerical comparison.