Muse Spark Open’s published one-million-token endpoint path uses eight NVIDIA H200 GPUs. At 141GB per device, the pool totals 1,128GB—428GB above Meta’s stated 700GB minimum for NVFP4 and 594.989GB above the official repository’s 533.011GB of listed safetensors. Hugging Face labels the named endpoint recipe prevalidated at 1,048,576-token context.
The long-context evidence
| Evidence | Value | Interpretation |
|---|---|---|
| Official listed NVFP4 safetensors | 533.011GB | Stored checkpoint size, not peak runtime memory |
| Meta’s stated NVFP4 minimum | 700GB | Published system-memory boundary |
| 8×H200 | 1,128GB nominal | 428GB above the stated minimum |
| Prevalidated endpoint context | 1,048,576 tokens | Specific to the named hosted recipe |
A specific Hopper exception, not a universal claim
Meta describes NVFP4 as targeting NVIDIA Blackwell GPUs. Hugging Face nevertheless publishes this specific H200 Hopper endpoint recipe. The careful conclusion is that the named hosted path is prevalidated—not that every self-managed Hopper runtime, container or topology supports the checkpoint.
Five H200 GPUs also clear 700GB arithmetically at 705GB, but the reviewed official sources do not validate that five-device configuration. Device count, placement, runtime and context are part of the evidence.
One million tokens still consume memory
The published context value is a recipe target, not a promise that every request can fill the window under every batch and concurrency setting. KV cache grows with sequence length and serving load. Multimodal processing, runtime allocations, communication buffers, uneven expert placement and temporary workspaces also consume memory.
The 428GB difference from Meta’s minimum averages 53.5GB per GPU only under even pressure. Measure each device: a deployment fails on the most constrained GPU, not on the aggregate average.
How the published endpoint choices differ
- 8×RTX PRO 6000: 768GB; prevalidated endpoint at 131,072-token context.
- 8×H200: 1,128GB; prevalidated endpoint at 1,048,576-token context.
- 4×GB300/B300-class: 1,152GB nominal; official self-managed SGLang TP4/EP4 reference.
- 8×B200: 1,440GB; official launch article’s single-node B200 path.
The extra H200 capacity does not by itself establish a speed ranking. Context, runtime, topology, batch, concurrency and implementation determine performance.
Use the complete evidence
The complete eight-H200 analysis includes exact margins, the Hopper/Blackwell evidence boundary, comparison table and nine-step reproduction checklist.
- Check Muse Spark Open’s exact VRAM requirements
- Compare every published hardware path
- Review H200 SXM and NVL specifications
- Stage a reproducible benchmark record
- Follow qualifying hardware listings
- Subscribe to the main RSS feed
- Read the latest Weekly Digest synthesis
Sources
- Official Muse Spark Open launch and endpoint recipes
- Official Muse Spark Open NVFP4 repository
- NVIDIA H200 specifications
Checked October 25, 2026. Prevalidation applies only to the named hosted recipe. Capacity and context support are not performance benchmarks.