Muse Spark Open now has two unusually clear hosted hardware lanes: Hugging Face publishes a prevalidated eight-GPU RTX PRO 6000 Blackwell recipe for 131,072-token context and an eight-H200 recipe for the 1,048,576-token tier. The choice is not simply 768GB versus 1,128GB; it is a different context target, hardware generation and validated configuration.
The published paths
| Path | Nominal aggregate memory | Context in published recipe | Evidence boundary |
|---|---|---|---|
| 8×RTX PRO 6000 Blackwell 96GB | 768GB | 131,072 | Prevalidated hosted endpoint recipe |
| 8×H200 141GB | 1,128GB | 1,048,576 | Prevalidated hosted endpoint recipe |
| 4×GB300/B300-class | 1,152GB nominal at 288GB tier | Workload-dependent | Official self-managed SGLang TP4/EP4 reference |
| 8×B200 180GB node | 1,440GB | Workload-dependent | Official article’s node-scale NVFP4 path |
NVIDIA documents RTX PRO 6000 Blackwell Server Edition at 96GB, H200 at 141GB, B200 at 180GB and the current B300 tier at 288GB. The Hugging Face launch article currently describes the RTX PRO endpoint devices as “98 GB.” HighVRAM calculates from NVIDIA’s 96GB specification and requires operators to record the memory their runtime exposes.
Why eight H200s are not merely “more VRAM”
The H200 configuration is tied to the published one-million-token endpoint recipe. Its additional 360GB of nominal aggregate memory versus eight 96GB RTX PRO cards is not a general speed ranking; context, cache, runtime implementation and topology all change the workload. Meta describes NVFP4 as Blackwell-oriented, while Hugging Face validates this specific H200 hosted path. Preserve that distinction.
Why four B200s remain an experiment
Four B200 GPUs total 720GB, just 20GB above Meta’s stated 700GB minimum. But the official article says NVFP4 runs on a single B200 node, and NVIDIA documents that node as eight 180GB GPUs. A four-device sum clears the arithmetic threshold; it does not inherit the eight-device reference’s topology, placement or validation.
Use the complete decision guide
The Muse Spark Open hardware guide compares all published and capacity-only paths, explains the 96GB/98GB source discrepancy and provides a buying and validation checklist.
- Check the exact 533.011GB files and 700GB minimum
- Compare NVIDIA high-memory tiers
- Stage a reproducible Muse Spark 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 Hugging Face/Meta launch and deployment recipes
- NVIDIA HGX hardware specifications
- NVIDIA supported-GPU memory table
Checked October 25, 2026. Capacity is not a performance benchmark. “Prevalidated” applies only to the named hosted recipe.