Scale Agentic Coding Workflows with Meta’s Muse Spark Open on NVIDIA GB300 NVL72

Meta and NVIDIA Target Agentic Coding with Muse Spark Open

Meta has released Muse Spark Open, a new 816-billion parameter Mixture-of-Experts (MoE) model specifically engineered for complex, agentic coding workflows. The model, optimized to run on NVIDIA's GB300 NVL72 platform, is distinguished by its ability to handle entire code repositories within its 1 million token context window. This release signals a significant move toward AI systems designed for sustained, multi-step software development tasks rather than simple, single-shot code completions.

Technical Specifications and Performance

Muse Spark Open utilizes a sparse MoE architecture that activates only 42 billion of its 816 billion total parameters for any given token, aiming to balance frontier capability with serving efficiency. The model is distributed with NVFP4 checkpoints, a 4-bit floating-point precision format that reduces the memory footprint and leverages the native Tensor Core acceleration of NVIDIA's Blackwell architecture. This tight integration with hardware allows the system to achieve high throughput for demanding agentic workloads.

• Total Parameters: 816B
• Active Parameters: 42B (MoE with top-8 routing across 256 experts)
• Context Window: 1M tokens
• Modality Support: Native processing of interleaved text, images, video, and audio
• Target Hardware: NVIDIA GB300 NVL72 Blackwell Ultra
• Performance: Up to 50K tokens per second per GPU; up to 600 tokens per second per user

Ecosystem Impact and Deployment

The collaboration extends beyond the model itself, encompassing a full software and deployment stack. NVIDIA is providing multiple pathways for developers, including a downloadable NVIDIA NIM for simplified production deployments and direct control via libraries like TensorRT-LLM. By offering a large, open model co-optimized with a specific hardware platform and a clear deployment path, the partnership provides enterprises a tangible route to implement sophisticated AI agents for tasks like codebase refactoring, debugging from screen recordings, and multi-file project planning.

The collaboration between Meta and NVIDIA on Muse Spark Open demonstrates that the frontier of enterprise AI is shifting from generalist models to specialized, hardware-aware systems. By co-optimizing an open model for a specific platform (GB300 NVL72) and a specific domain (agentic coding), they provide a high-performance, production-ready path that lowers the barrier to entry for complex, long-running AI workflows.