Muse Spark Open’s official four-device self-managed reference uses SGLang across four GB300 GPUs. The published command sets tensor parallelism 4 and expert parallelism 4 for the NVFP4 checkpoint. Using NVIDIA’s documented 288GB B300 memory tier for the GB300 platform, the four-device pool totals 1,152GB—452GB above Meta’s stated 700GB minimum.
The four-device reference
| Evidence | Value | Interpretation |
|---|---|---|
| Official listed NVFP4 safetensors | 533.011GB | Stored checkpoint size, not peak runtime memory |
| Meta’s stated NVFP4 minimum | 700GB | Published memory boundary |
| 4×B300 288GB tier | 1,152GB nominal | 452GB above the stated minimum |
| Published parallelism | TP4 + EP4 | All four devices participate in both dimensions |
Why the GB300/B300 wording needs care
The launch article says the NVFP4 reference runs across “four GB300 GPUs.” NVIDIA describes GB300 systems as platforms built around B300 GPUs and documents a 288GB-per-GPU tier. HighVRAM preserves the article’s wording while using the documented B300 memory figure. Operators should record the exact system SKU and runtime-visible memory rather than treating the platform label as a complete specification.
The command is the useful evidence
The official SGLang recipe names the NVFP4 checkpoint, modelopt_fp4 quantization, the flashinfer_trtllm mixture-of-experts backend, --tp-size 4 and --ep-size 4. That is stronger evidence than adding four GPU capacities and assuming the software will follow.
It does not publish a universal speed, power, context, reliability or concurrency result. Reproduction still depends on system topology, firmware, drivers, CUDA, integration version, placement, context, modality and serving load.
Why this is not the same as four B200s
Four B200 GPUs total 720GB and clear the 700GB minimum by 20GB, but no reviewed official source validates that four-device B200 configuration. Four GB300-class devices are explicitly named in the SGLang reference and have much more nominal memory headroom. The distinction is published deployment evidence—not merely 1,152GB versus 720GB.
Use the complete reference analysis
The complete four-GB300 guide includes the exact SGLang parameters, naming disclosure, path comparison and nine-step reproduction checklist.
- Check Muse Spark Open’s exact VRAM requirements
- Compare every published hardware path
- Review NVIDIA B300 and other high-memory tiers
- 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 SGLang reference
- Official Muse Spark Open NVFP4 repository
- NVIDIA HGX GB300 and B300 specifications
Checked October 25, 2026. Nominal capacity is not a performance benchmark. The reference applies to the named SGLang configuration and hardware class.