MiniMax H3 disk space: 108 GB for SGLang, 19.5 GiB for INT8

Plan for about 108 GB of local disk if you serve MiniMax H3 with SGLang, which is the figure in MiniMax's self-hosting guide. A ComfyUI install can be far smaller, because it uses pruned and quantized files: the integration index lists 19.53 GiB for INT8 ConvRot and 61.73 GiB for the BF16 original.

Why do three numbers describe one model?

They answer different questions, and they use different units. The guide's 108 GB is a reference-setup budget for the official mixed BF16/FP32 weights. The index's sizes are per safetensors or GGUF file, in GiB. A GiB is about 7 percent larger than a GB, so 61.73 GiB is about 66.3 GB.

What else needs room?

The README's download command fetches model_index.json plus both FL2VA/* and Ref2VA/*, while the guide's quickstart downloads only FL2VA/*. The two variants are separate partitions, and each runs as its own SGLang service, so budget for both if you want text-to-video and reference-to-video. Add space for your outputs, logs and any second quantization you test; the guide says not to mix quantized and lossless outputs in benchmarks, which usually means keeping both.

How should you budget for a team?

Count each copy. A shared network volume holds one set; each GPU node that loads the model needs fast access to it, and the SGLang route wants the weights on local disk. Two nodes with both variants is two times the 108 GB figure at least, before outputs and logs. The guide's number is for the reference setup, so a pruned ComfyUI install on a workstation will be much smaller.

Slow disks also show up as slow starts. A model of this size takes time to load from a network drive, which you pay for on every restart. The sources I read give no load-time figure, so measure it.

What does a hosted job need?

No weights on disk. A Sume minimax-h3 clip comes back as an MP4 you download from unsigned_urls[0]. The video docs list the fields.