Four NVIDIA B200 GPUs pass Muse Spark Open’s published VRAM minimum on paper—but only narrowly. Each B200 provides 180GB of HBM3e, for 720GB aggregate. That is 20GB above Meta’s stated 700GB minimum for the NVFP4 model and 186.989GB above the official repository’s 533.011GB of listed safetensors. It is not, however, an officially validated four-GPU deployment.
The four numbers that matter
| Evidence | Value | What it proves |
|---|---|---|
| Official listed NVFP4 safetensors | 533.011GB | Stored checkpoint size, not peak runtime memory |
| Meta’s stated NVFP4 minimum | 700GB VRAM | Published system-memory boundary |
| Four B200 GPUs | 720GB nominal | Aggregate arithmetic clears the minimum by 20GB |
| Published B200 node path | 8×180GB = 1,440GB | The official node-scale B200 configuration, not evidence for four GPUs |
Why the 20GB margin is not a deployment guarantee
If model state divided perfectly evenly, the difference between 720GB and the stated minimum would average only 5GB per GPU. Real placement need not divide evenly: attention layers, shared experts, routed experts, the vision stack, projector and output head can create different peaks on different devices. CUDA and framework allocations, communication buffers, KV cache, context length, batch size and concurrency also consume memory.
Aggregate capacity therefore answers only the first question. A valid four-B200 result still needs exact hardware identity, exposed memory per device, topology, runtime and kernel versions, tensor/expert placement, offload state, context, concurrency, per-device peak memory, correctness and failure behavior.
What the official evidence actually validates
The official Muse Spark Open launch article says the NVFP4 model runs on a single B200 node. NVIDIA documents the current HGX B200 node as eight B200 GPUs with 180GB each. The published self-managed SGLang reference instead uses four GB300/B300-class GPUs with TP4 and EP4. Neither source establishes a four-B200 reference configuration.
That makes 4×B200 a useful experiment for an operator who already controls the hardware—not a configuration to buy on the strength of capacity arithmetic alone.
Read, test or follow the evidence
Use the complete four-B200 analysis for the comparison table and nine-step validation plan.
- Check the exact Muse Spark Open VRAM requirements
- Compare published and capacity-only hardware paths
- Review the B200 180GB capacity guide
- Stage a reproducible benchmark record
- Follow high-VRAM hardware listings
- Subscribe to the main RSS feed
Need a consumer-class Muse model instead? Muse Glimmer 30B is a different model family with official 24GB Q4_K_M and 32GB dynamic Q4 targets. Use the Muse Glimmer hardware guide; do not transfer Spark’s capabilities, artifacts or benchmark results to Glimmer.
\n\nSources
- Official Muse Spark Open NVFP4 repository
- Official Muse Spark Open launch and deployment article
- NVIDIA HGX B200 and GB300 platform specifications
Checked October 25, 2026. Nominal aggregate memory is not proof of runtime compatibility, model correctness or performance.