XNews

Meta Launches Muse Spark Open 816B Coding Model Trained on NVIDIA GB300 NVL72

Meta released Muse Spark Open, an 816B mixture-of-experts coding model with 42B active parameters and a 1M-token context window, trained on NVIDIA GB300 NVL72 Blackwell Ultra. Meta produced NVFP4 checkpoints optimized for Blackwell Ultra, delivering over 50K tokens per second per GPU and up to 600 tokens per second per user. The model is available on Hugging Face with NVIDIA NIM for production deployments.

Event version 1.

Key insights

Mixture-of-experts architecture

Muse Spark Open activates 42B of 816B total parameters per token through top-8 routing across 256 experts. The model is designed for agentic coding workflows with higher first-attempt accuracy, reliable tool calling, and deep codebase understanding.

1M token context window

The 1M token context window holds an entire project including complex dependencies, legacy code, and thousands of files without breaking tasks into pieces. The model supports multistep task execution with tool use, output evaluation, and iterative refinement.

Multimodal perception

Muse Spark Open adds audio and video perception alongside image and document understanding. It processes interleaved modalities natively through a 1.8B perception encoder.

NVFP4 optimization for Blackwell Ultra

NVFP4 checkpoints generated by Meta and optimized for NVIDIA Blackwell Ultra shrink the memory footprint of all 816B parameters with native FP4 Tensor Core acceleration. Running agentic workloads on NVIDIA GB300 NVL72 delivers up to 50K tokens per second per GPU and up to 600 tokens per second per user.

Availability

Muse Spark Open NVFP4 weights are available for download from Hugging Face. NVIDIA NIM is available for production deployments with a single-command setup.

Source posts