Scale Agentic Coding Workflows with Meta’s Muse Spark Open on NVIDIA GB300 NVL72
Oct 15, 2026, 11:00 AM · NVIDIA Developer
NVIDIA’s serving note: Muse Spark Open trained on GB300 NVL72, NVFP4 for Blackwell, claims up to 50K tok/s per GPU for agentic coding.
Why it matters
NVIDIA’s developer blog says Meta’s Muse Spark Open — 816B MoE with 42B active, 1M context — was trained on NVIDIA GB300 NVL72 Blackwell Ultra for agentic coding. NVFP4 checkpoints optimized for Blackwell shrink the memory footprint with native FP4 Tensor Core acceleration. NVIDIA claims running agentic workloads on GB300 NVL72 delivers up to 50K tokens per second per GPU and up to 600 tokens per second per user. The model adds audio and video perception via a 1.8B perception encoder alongside image and document understanding.
From the desk
This is the co-marketing layer on Meta’s open release: train on our rack, serve on our FP4 path, download NIM. For teams already on Blackwell, it’s a concrete deployment story. For everyone else, it’s a reminder that “open weights” still have a preferred silicon.
The agentic-coding framing is right — million-token repos and tool calling need sustained throughput, not just single-shot codegen. We’ll want independent tokens-per-second numbers under real harnesses.
I’m watching NIM one-command deploys versus raw Hugging Face NVFP4, and whether 600 tok/s per user shows up outside NVIDIA’s blog.
Context
Companion Meta and Hugging Face posts detail architecture; this piece focuses on GB300 training and Blackwell serving.
Who feels it
- Platform engineers
- A Blackwell-first path to host Muse Spark Open for coding agents.
- Meta open-weight users
- Clear NVFP4 + NIM options beside BF16 clusters.
- Competing chip vendors
- Pressure to match MoE serving efficiency narratives.
What to watch
- Independent GB300 throughput measurements
- NIM adoption versus community vLLM/SGLang recipes
- Multi-user long-context KV cache behavior in production