All local H3 guides
Hardware / Local H3

MiniMax H3 hardware requirements

MiniMax H3 VRAM, system RAM, storage, and GPU requirements for ComfyUI, dual RTX 5090, DGX Spark, H100, H200, B300, and AMD Instinct.

Reviewed
2026-08-09
Reading time
8 min
Evidence
2 official / 3 community
Short answer

There is no single official minimum GPU for every H3 runtime. Original BF16 serving is a multi-GPU workload; current SGLang recipes use four H100/H200-class GPUs or eight B300-class GPUs for resident serving. Its verified consumer recipe uses two RTX 5090 cards with layerwise offload. Community ComfyUI repacks can run on much less VRAM, including a reported 16 GB laptop GPU, but require heavy model swapping and are not equivalent to the original BF16 deployment.

01

Why the requirement changes by runtime

H3 combines a 33B dense omni-transformer, the full Qwen3-VL-32B encoder, a visual VAE, and an audio VAE. Keeping every component resident needs far more accelerator memory than loading the encoder once, moving it away, and streaming transformer or decoder blocks from system memory. Quantization reduces weight memory again, but it changes the precision path.

  • Resolution and frame count grow activation memory.
  • Ref2VA adds reference tokens and can be heavier than text-only generation.
  • Offloading lowers VRAM but increases PCIe or unified-memory traffic.
  • Downloading only FL2VA or Ref2VA reduces disk use, not the memory of a loaded task.
02

Measured configurations, not universal minimums

ConfigurationMeasured profileReported peak
4 x H100SGLang, BF16/FP32, 5 s at 1344 x 76849.80 to 66.04 GB per GPU
4 x H200SGLang, BF16/FP32, 5 s at 1344 x 76863,490 to 94,290 MB per GPU
8 x B300SGLang BF16 or online FP8about 52 to 124 GB per GPU by profile
2 x RTX 5090SGLang layerwise offload, 5 s at 1344 x 76826.3 GiB per GPU
1 x DGX SparkCommunity Diffusers INT8 plus group offloadFits 128 GB unified memory
RTX 4090 Laptop 16 GBCommunity ComfyUI repack, 5 s at 960 x 540Dynamic loading; 16 GB VRAM
03

Plan system RAM and storage with VRAM

Low-VRAM workflows rely on system memory. A machine that technically loads the diffusion model can still stall, swap, or fail during text encoding and video decode. The L40S community benchmark reported 64 GB system RAM as comfortable for its optimized raw-weight path. The DGX Spark guide allocated roughly 144 GB to selected FL2VA files and advised 180 GiB of free storage.

  • Use fast local NVMe for model caches and outputs.
  • Keep at least tens of gigabytes beyond selected model files for containers, caches, and videos.
  • Avoid downloading Ref2VA when the machine will only serve text and keyframe jobs.
  • Measure peak host RAM through a complete decode, not only model load.
04

What to buy for each goal

Do not choose hardware from parameter count alone. Benchmark the exact task family, duration, resolution, references, steps, precision, attention backend, and runtime that will ship.

GoalPractical hardware direction
Explore locallyA high-memory NVIDIA workstation with ComfyUI repacks and patience
Preserve original BF16 pathUse a current measured multi-GPU SGLang recipe
Serve several usersDatacenter GPUs, fast interconnect, and measured batching
Test one exact machineRent that SKU before purchasing and run a fixed canary
Frequently asked questions
How much VRAM does MiniMax H3 need?

It depends on checkpoint precision, resolution, runtime, and offloading. SGLang measured roughly 50 to 94 GB peak per GPU on four-card datacenter recipes and 26.3 GiB per GPU on its verified dual RTX 5090 offload recipe. Quantized ComfyUI repacks can run with less.

Can an RTX 4090 run MiniMax H3?

A community report generated a 960 x 540 five-second clip on a 16 GB RTX 4090 Laptop GPU using pruned INT8 diffusion weights, a quantized encoder, and dynamic loading. Treat that as evidence of feasibility for that workflow, not an official minimum or a desktop 4090 guarantee.

How much disk space does MiniMax H3 use?

Storage depends on whether you download FL2VA, Ref2VA, both task families, original BF16 files, or smaller ComfyUI repacks. One documented DGX Spark FL2VA setup required about 144 GB of selected H3 files and recommended at least 180 GiB free before container images and outputs.

Does MiniMax H3 run on AMD GPUs?

SGLang publishes measured resident recipes for MI300X and MI355X, including one-, two-, four-, and eight-GPU scaling runs. Consumer AMD support is a different question and is not established by those datacenter results.

Sources and evidence

Read the original material

Community figures describe the named machine and workflow only. They are not MiniMax performance guarantees.

  1. Official documentationSGLang H3 deployment and benchmarks
  2. Official documentationMiniMax H3 architecture and checkpoints
  3. Community benchmarkDGX Spark INT8 and offload report
  4. Community benchmarkRTX 4090 Laptop ComfyUI report
  5. Community benchmarkL40S raw-weight optimization report