- Published on
Scale Agentic Coding Workflows with Meta’s Muse Spark Open on NVIDIA GB300 NVL72
- Authors
- Name
- DeepHorus
Meta has launched Muse Spark Open, an 816-billion-parameter Mixture-of-Experts model with 42 billion active parameters and a 1-million-token context window. It was trained on NVIDIA GB300 NVL72 Blackwell Ultra systems for agentic coding workflows. The model targets sustained multistep work rather than single-shot code completions.
General coding models excel at discrete tasks like writing functions or unit tests. Agentic coding requires planning multifile refactors, navigating thousands of files, recovering from failed tests, and coordinating subagents. Meta designed Muse Spark Open to plan, act, observe, and self-correct over long horizons.
The model uses sparse MoE routing that activates 42B of 816B parameters per token, sending each token to eight of 256 experts. It includes a 1.8B perception encoder for interleaved text, images, video frames, and audio. NVFP4 checkpoints shrink memory footprint with native FP4 Tensor Core acceleration.
A 1M-token context window holds entire projects, including complex dependencies, legacy code, and thousands of files without breaking tasks into pieces. The model also supports web search, subagent spawning, and computer use across desktop, mobile, and browser. It generates separate commentary before each action so developers understand what the agent is doing and why.
The GB300 NVL72 integrates 72 Blackwell Ultra GPUs in a rack-scale platform. Its 72-GPU NVLink domain delivers 130 TB/s, eliminating bottlenecks for expert traffic. That bandwidth determines how much theoretical throughput Muse Spark Open deployments realize.
Meta says the model delivers higher first-attempt accuracy, reliable tool calling, and deep codebase understanding. Freed memory supports larger KV cache for concurrent long-context sessions, increasing tokens per second per user and lowering cost per token. That lowers serving costs.
The launch signals a shift toward agentic coding systems for sustained, multistep work. Muse Spark Open on GB300 NVL72 delivers improvements out of the box, Meta said. Adoption across production workflows remains uncertain.
