Short answer: yes—the official Muse Spark Open launch article publishes an SGLang NVFP4 reference configuration across four GB300 GPUs, with tensor parallelism 4 and expert parallelism 4. Using NVIDIA’s documented 288GB B300 GPU memory tier for the GB300 platform, four devices provide 1,152GB nominal aggregate memory: 452GB above Meta’s stated 700GB NVFP4 minimum and 618.989GB above the official repository’s 533.011GB of listed safetensors.
Naming matters
The launch article says “four GB300 GPUs.” NVIDIA’s platform documentation describes GB300 systems built around B300 GPUs, with 288GB per GPU in the referenced tier. HighVRAM preserves the article’s wording while using the documented B300 memory figure for arithmetic. Confirm the exact system SKU and runtime-visible memory rather than treating a platform name as a complete bill of materials.
Four GB300-class GPUs: the exact boundary
| Boundary | Value | Margin in the 4×288GB pool |
|---|---|---|
| Official listed NVFP4 safetensors | 533.011GB | 618.989GB above stored files |
| Meta’s stated NVFP4 minimum | 700GB VRAM | 452GB above the minimum |
| 4×B300 288GB tier | 1,152GB nominal | 288GB per device |
| Published parallelism | TP4 + EP4 | All four devices participate in both dimensions |
What the official SGLang recipe actually specifies
- The official NVFP4 checkpoint:
meta-models/Muse-Spark-Open-816B-A42B-NVFP4 - SGLang’s Meta integration environment and
sglang_meta serve modelopt_fp4quantization- The
flashinfer_trtllmmixture-of-experts runner backend --tp-size 4and--ep-size 4- Text, image, audio and video input support in the integration
The recipe does not publish a universal throughput, latency, wall-power or maximum-concurrency result. It also does not make the full one-million-token architectural context free: KV cache and serving state still scale with the chosen workload.
Why this is stronger than four-B200 arithmetic
Four B200 GPUs total 720GB and clear Meta’s minimum by only 20GB, but the reviewed official sources do not validate that four-device B200 configuration. The four-GB300 route is different: the official launch article gives a concrete SGLang command and explicitly sets TP4 and EP4.
That does not make every four-device B300 system interchangeable. Firmware, drivers, CUDA, container contents, interconnect, exposed memory and the integration revision remain part of a reproducible deployment.
GB300 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 | Lower-capacity hosted path |
| 8×H200 141GB | 1,128GB | Prevalidated hosted endpoint at 1,048,576 context | Long-context hosted path |
| 4×GB300/B300-class 288GB | 1,152GB nominal | Official SGLang TP4/EP4 command | Self-managed four-device reference |
| 8×B200 180GB node | 1,440GB | Official launch article’s single-node B200 path | Node-scale Blackwell path |
Reproduce it without erasing the evidence boundary
- Record the exact GB300 system and B300 GPU SKU, firmware and visible memory.
- Document the full NVLink/NVSwitch and host topology.
- Pin the checkpoint revision plus the SGLang integration, container, driver and CUDA versions.
- Preserve
modelopt_fp4, the FlashInfer/TensorRT-LLM MoE backend, TP4 and EP4 unless testing a named variation. - Disclose any CPU or storage offload.
- Fix context, modality, batch size and concurrency.
- Capture idle, load and peak memory on every device.
- Check output correctness, retries, errors and out-of-memory behavior.
- Report latency, throughput, warmup, repetitions, throttling and wall power separately.
Decision
Choose the four-GB300 SGLang route when reproducing the official self-managed NVFP4 reference is the priority. Choose the eight-RTX-PRO or eight-H200 recipes when a prevalidated hosted endpoint is preferable. Do not substitute four B200s merely because their 720GB sum crosses 700GB; that is a capacity experiment, not the published four-device reference.
Reproducing this path? Follow the official Muse Spark Open runtime map for the current SGLang and vLLM four-B300 recipes, command boundaries and smoke-test sequence.
Continue the research
- See the exact Muse Spark Open VRAM requirements
- Compare every published and capacity-only path
- Review NVIDIA B300 and other high-memory tiers
- Contrast the unvalidated four-B200 capacity path
- Stage a reproducible Muse Spark record
- Follow qualifying high-VRAM hardware listings
Primary 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 memory is not a performance benchmark. The official reference applies to the named SGLang configuration and documented hardware class.