Meta’s new Muse Spark Open creates a useful hardware trap: its official NVFP4 repository lists 533.011GB of safetensors files, but Meta says a self-hosted deployment needs a GPU system with at least 700GB of VRAM. A pool can pass the download-size test and still fail the publisher’s runtime boundary.
The capacity gap
| Configuration | Aggregate VRAM | After 533.011GB files | Against 700GB minimum |
|---|---|---|---|
| 8×80GB | 640GB | +106.989GB | 60GB short |
| 5×H200 141GB | 705GB | +171.989GB | 5GB above |
| 4×B200 180GB | 720GB | +186.989GB | 20GB above |
| 8×96GB | 768GB | +234.989GB | 68GB above |
The 640GB row is the caution. Eight 80GB accelerators can hold the listed files in aggregate, yet the pool remains below Meta’s minimum. Five H200s cross the number by only 5GB, and format support is a separate gate: Meta positions the NVFP4 checkpoint for NVIDIA Blackwell hardware.
Why the files are not the deployment
A live server needs more than stored weights: runtime allocations, execution representations, kernels, communication buffers, KV cache, multimodal processing and imperfect per-device placement all consume memory. Muse Spark Open is an 816B-total, roughly 42B-active mixture-of-experts model, but “active” parameters describe token-time compute—not how little of the model can be stored.
The advertised one-million-token context is another ceiling, not a free operating point. Context, batch and concurrency must be raised only after measuring peak memory on every device.
Use the full evidence guide
The Muse Spark Open VRAM requirements guide documents all 127 official weight shards, the exact repository revision, capacity tiers through 864GB, Blackwell-format caveats and a deployment checklist.
- Compare NVIDIA high-memory tiers
- Plan explicit multi-GPU placement
- Contribute a reproducible measurement
- 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
Checked October 25, 2026. File totals use decimal GB. Aggregate capacity does not prove a supported or successful deployment.