Reproducing the official self-managed path? See whether four GB300 GPUs can run Muse Spark Open, including the published SGLang TP4/EP4 configuration and GB300/B300 naming boundary.
Evaluating the published long-context path? See whether eight H200 GPUs can run Muse Spark Open, including the prevalidated 1,048,576-token endpoint recipe and its Hopper-specific evidence limits.
Short answer: yes—there is a published, prevalidated Muse Spark Open endpoint recipe using eight RTX PRO 6000 Blackwell GPUs. NVIDIA specifies 96GB per Server Edition GPU, so the eight-device pool is 768GB: 68GB above Meta’s stated 700GB NVFP4 minimum and 234.989GB above the official repository’s 533.011GB of listed safetensors. That evidence applies to the named recipe and its 131,072-token context setting; it is not a promise that any arbitrary eight-card workstation or server will work.
Evidence boundary
Published path: eight RTX PRO 6000 Blackwell devices in Hugging Face’s prevalidated endpoint recipe. Not established: every chassis, PCIe topology, driver stack, cooling design or self-managed runtime that happens to contain eight cards.
Eight RTX PRO 6000 GPUs: the exact arithmetic
| Boundary | Value | Margin in the 8×96GB pool |
|---|---|---|
| Official listed NVFP4 safetensors | 533.011GB | 234.989GB above stored files |
| Meta’s stated NVFP4 minimum | 700GB VRAM | 68GB above the minimum |
| 8×RTX PRO 6000 Blackwell | 768GB nominal | 96GB per device |
| Published endpoint context | 131,072 tokens | Recipe-specific, not a universal ceiling |
The launch article currently describes the endpoint GPUs as “98 GB,” while NVIDIA’s product specification lists 96GB of GDDR7 with ECC. HighVRAM uses the manufacturer’s 96GB figure for arithmetic and recommends recording the exact memory exposed by the runtime. Eight 96GB devices total 768GB; eight hypothetical 98GB devices would total 784GB, but smoothing over that source discrepancy would make the buying guidance less reliable.
Why this path is stronger than a capacity-only guess
A prevalidated recipe establishes more than an aggregate sum. It shows that a named model artifact, hardware family, device count, runtime path and context target have been assembled together successfully by the publishing platform. By contrast, four B200 GPUs also clear 700GB at 720GB aggregate, but the reviewed official material does not validate a four-B200 Muse Spark deployment.
Prevalidation still does not establish your throughput, latency, power draw or maximum reliable concurrency. Those require measurements from the exact deployment.
What the 68GB margin must still hold
- CUDA, driver and framework allocations
- Tensor-parallel and expert-parallel communication buffers
- Uneven placement of attention, routed experts, shared experts, vision components, projector and output head
- KV cache for the chosen context, batch size and concurrency
- Input preprocessing and multimodal state
- Temporary workspace memory and fragmentation
The 68GB aggregate difference from Meta’s minimum averages 8.5GB per device only if pressure divides evenly. A run succeeds or fails on the most constrained device, not on the average.
RTX PRO 6000 versus the other published paths
| Path | Nominal memory | Published evidence | Best interpretation |
|---|---|---|---|
| 8×RTX PRO 6000 96GB | 768GB | Prevalidated hosted endpoint at 131,072 context | Lowest-capacity published recipe reviewed here |
| 8×H200 141GB | 1,128GB | Prevalidated hosted endpoint at 1,048,576 context | Published long-context hosted path |
| 4×GB300/B300-class 288GB | 1,152GB nominal | Official SGLang TP4/EP4 self-managed reference | Published four-device Blackwell reference |
| 8×B200 180GB node | 1,440GB | Official launch article’s single-node B200 path | Published node-scale B200 route |
Validate before buying or reproducing it
- Confirm the exact RTX PRO 6000 edition and runtime-visible memory on every device.
- Record the chassis, slot layout, PCIe generation and peer-to-peer topology.
- Pin the Muse Spark repository revision, quantization and all runtime/container versions.
- Record tensor-parallel and expert-parallel sizes plus the complete placement map.
- State whether weights, KV cache or other tensors are offloaded to host memory.
- Fix context, modality, batch size and concurrency before comparing results.
- Capture per-device idle, load and peak memory rather than aggregate memory alone.
- Check output correctness, errors, retries and out-of-memory failures.
- Report latency, throughput, warmup, repetitions, thermals and wall power separately.
Decision
Use the eight-RTX-PRO route when you want the lowest-capacity published Muse Spark recipe reviewed here and 131,072-token context matches the job. Prefer the eight-H200 hosted recipe when the published one-million-token target matters, or the four-GB300 SGLang reference when reproducing the official self-managed path is the priority. Do not treat the RTX PRO recipe as a benchmark or as validation of an arbitrary custom chassis.
Continue the research
- See the exact Muse Spark Open VRAM requirements
- Compare every published and capacity-only hardware path
- Review RTX PRO 6000 Blackwell specifications and fit
- Understand multi-GPU placement limits
- Stage a reproducible Muse Spark benchmark record
- Follow qualifying high-VRAM hardware listings
Primary sources
- Official Muse Spark Open launch and prevalidated endpoint recipes
- Official Muse Spark Open NVFP4 repository
- NVIDIA RTX PRO 6000 Blackwell specifications
- NVIDIA HGX B200 and GB300 platform specifications
Checked October 25, 2026. “Prevalidated” applies only to the named hosted recipe. Capacity is not a performance benchmark or a purchase guarantee.