AI

Meta Launches Muse Spark Open Coding Model on NVIDIA Blackwell

Thursday, October 15, 2026Read Original

Details

  • Meta released Muse Spark Open, an open-weight multimodal mixture-of-experts model designed for coding and agentic workloads.
  • The model has 816 billion total parameters, with 42 billion activated for each token, and supports a context window of up to 1 million tokens.
  • Muse Spark Open accepts text, images, audio, and video inputs, enabling agents to work across large repositories and multimodal project materials.
  • Meta trained the model on NVIDIA GB300 NVL72 systems based on the Blackwell Ultra platform.
  • Meta also released BF16 and NVFP4 checkpoints, with NVFP4 using 4-bit precision to reduce memory requirements on compatible NVIDIA hardware.
  • NVIDIA says the optimized configuration can exceed 50,000 tokens per second per GPU and reach up to 600 tokens per second per user.
  • The sparse architecture activates only part of the model for each token, aiming to reduce serving costs compared with a dense model of similar total size.
  • Developers can download the weights or deploy the model through supported inference runtimes and hosted providers.

Impact

The release raises the hardware and software bar for open-weight coding models while making million-token context and multimodal inputs available for self-hosted or hosted deployments. Its practical reach will depend on access to high-end Blackwell or similarly capable systems, so the model expands openness without eliminating infrastructure barriers. The NVIDIA optimization also ties adoption of Meta’s model to Blackwell-based AI infrastructure and positions agentic coding as a major workload for next-generation data centers.

Rift Dispatch