The lowest-capacity published Muse Spark Open recipe reviewed by HighVRAM uses eight RTX PRO 6000 Blackwell GPUs. NVIDIA specifies 96GB per Server Edition device, so the pool totals 768GB—68GB above Meta’s stated 700GB minimum for the NVFP4 model. Hugging Face identifies the eight-device endpoint recipe as prevalidated at 131,072-token context.
The published boundary
| Evidence | Value | Interpretation |
|---|---|---|
| Official listed NVFP4 safetensors | 533.011GB | Stored checkpoint size, not peak runtime memory |
| Meta’s stated NVFP4 minimum | 700GB | Published deployment boundary |
| 8×RTX PRO 6000 Server Edition | 768GB nominal | 68GB above the stated minimum |
| Prevalidated endpoint context | 131,072 tokens | Specific to the published recipe |
A source discrepancy worth preserving
The launch article currently describes the endpoint devices as “98 GB.” NVIDIA’s Server Edition specification lists 96GB GDDR7 with ECC. HighVRAM uses the manufacturer’s 96GB figure, producing the conservative 768GB total, and recommends recording the exact memory exposed by the runtime.
That discrepancy does not erase the published recipe. It does mean buyers should avoid silently converting the launch prose into a product specification.
What prevalidated does—and does not—mean
The named recipe is stronger evidence than simply adding GPU capacities: it binds a hardware family, device count, model artifact, runtime path and context target. It does not validate every custom eight-card chassis, nor does it promise a particular latency, throughput, power draw or concurrency level.
The 68GB aggregate margin averages 8.5GB per GPU only under perfectly even pressure. Placement of attention, experts, vision components and the output head can be uneven, while framework state, communication buffers, KV cache and temporary workspaces consume additional memory.
How it compares
- 8×RTX PRO 6000: 768GB; prevalidated hosted recipe at 131,072 context.
- 8×H200: 1,128GB; prevalidated hosted recipe at 1,048,576 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.
Use the complete evidence
The complete eight-RTX-PRO analysis includes the exact margins, source discrepancy, hardware comparison and nine-step reproduction checklist.
- Check Muse Spark Open’s exact VRAM requirements
- Compare every published hardware path
- Review RTX PRO 6000 Blackwell specifications
- Stage a reproducible benchmark record
- Follow qualifying hardware listings
- Subscribe to the main RSS feed
Sources
- Official Muse Spark Open launch and endpoint recipes
- Official Muse Spark Open NVFP4 repository
- NVIDIA RTX PRO 6000 Blackwell Server Edition specifications
Checked October 25, 2026. Prevalidation applies only to the named recipe. Capacity is not a performance benchmark or a purchase guarantee.