Running MiniMax H3 Locally: ComfyUI Setup, Faster Downloads, and VRAM Peak Management

Scope and Preconditions

RTX 4090 with 24 GB VRAM, 64 GB system RAM, Windows, isolated ComfyUI 0.30.0 environment. The pipeline can generate video with native audio and output MP4, but a simple "download‑four‑files → open workflow → click Queue" is not production‑ready.

Build an Isolated Environment

mkdir H3-ComfyUI
cd H3-ComfyUI

git clone https://github.com/Comfy-Org/ComfyUI.git tools/comfyui
py -3.12 -m venv .venv-comfy

.\.venv-comfy\Scripts\python.exe -m pip install \
  torch==2.9.1 torchvision==0.24.1 torchaudio==2.9.1 \
  --index-url https://download.pytorch.org/whl/cu130

.\.venv-comfy\Scripts\python.exe -m pip install -r .\tools\comfyui\requirements.txt

Pin ComfyUI to a verified commit that contains an H3 VAE device‑conversion fix:

git -C .\tools\comfyui checkout 16e3f3034f2bba1fff6c70cbd759339778555cd6

Directory layout:

H3-ComfyUI/
├─ .venv-comfy/
└─ tools/
   └─ comfyui/
      ├─ models/diffusion_models/
      ├─ models/text_encoders/
      ├─ models/vae/
      └─ user/default/workflows/

First Hurdle: 42.5 GB Model Download

The four required files total about 42.5 GB (≈39.6 GiB on Windows):

FL2VA INT8 ConvRot DiT – 20.97 GB – models/diffusion_models/ Qwen3‑VL 32B NVFP4/AWQ – 15.69 GB – models/text_encoders/ Video VAE FP16 – 5.21 GB – models/vae/ Audio VAE FP32 – 0.61 GB – models/vae/ For GPUs with <12 GB VRAM the author recommends first testing T2V/I2V at 480 p before attempting the full 63 GB weight set.

Accelerating the Download

Downloading from Hugging Face on a Chinese network yielded ~150 KiB/s (≈38 h). Switching to ModelScope raised speeds to 92 MiB/s (Video VAE) and 86 MiB/s (Audio VAE). The following aria2c script uses 16 connections, disables file allocation, and enables resume:

$base = 'https://modelscope.cn/models/Comfy-Org/MiniMax-H3/resolve/master'
$root = '.\tools\comfyui\models'

aria2c -c -x16 -s16 -k8M --file-allocation=none \
  -d "$root\diffusion_models" \
  -o 'minimax_h3_fl2va_pruned_int8_convrot.safetensors' \
  "$base/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors"

aria2c -c -x16 -s16 -k8M --file-allocation=none \
  -d "$root\text_encoders" \
  -o 'qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors' \
  "$base/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors"

aria2c -c -x16 -s16 -k8M --file-allocation=none \
  -d "$root\vae" \
  -o 'minimax_h3_video_vae_fp16.safetensors' \
  "$base/vae/minimax_h3_video_vae_fp16.safetensors"

aria2c -c -x16 -s16 -k8M --file-allocation=none \
  -d "$root\vae" \
  -o 'minimax_h3_audio_vae_fp32.safetensors' \
  "$base/vae/minimax_h3_audio_vae_fp32.safetensors"

After download, verify SHA256 hashes (provided in the article) to avoid silent corruption. Example PowerShell verification:

Get-FileHash .\tools\comfyui\models\vae\minimax_h3_video_vae_fp16.safetensors -Algorithm SHA256

Run the Official Workflow First

.\.venv-comfy\Scripts\python.exe .\tools\comfyui\main.py \
  --listen 127.0.0.1 \
  --port 8189 \
  --reserve-vram 4

Open http://127.0.0.1:8189, select the MiniMax H3 T2V or I2V template, use a small canvas (e.g., 864×480) and 20 steps for a quick probe. Smoke‑test sequence:

864×480 – ~5.17 s (≈124 frames at 24 fps)

1056×608 or 608×1056 – ~5.17 s

Confirm video and audio are written to MP4

Scale up to 0.7 MP, 0.8 MP

H3 outputs at a fixed 24 fps; a 5‑second request yields 124 frames because of the 17k+5 frame grid constraint.

SageAttention Acceleration (Version‑Matched)

Official docs claim SageAttention roughly doubles H3 generation speed for the attention/sampling portion, with minimal quality loss. It does not affect model loading, text encoding, VAE decoding, or MP4 packaging.

Windows stack used:

Triton Windows 3.5.1.post24

SageAttention 2.2.0 cu130 wheel for torch 2.9+

KJNodes 1.4.9 at commit

6edfa76599c13905968841b81309784a7dbb7803

.\.venv-comfy\Scripts\python.exe -m pip install triton-windows==3.5.1.post24

Install the matching SageAttention wheel (downloaded from woct0rdho/SageAttention Windows releases) and KJNodes via git checkout, then restart ComfyUI and connect the SageAttention patch node to the UNETLoader.

Occasional fallback messages to PyTorch attention are expected and harmless.

EasyCache: Draft‑Only, Not Full‑Speed

Configuration for quick previews:

reuse_threshold = 0.20
start_percent   = 0.15
end_percent     = 0.95

In repeated 20‑step tests EasyCache skipped 6‑11 steps, reporting speed‑ups between 1.43× and 2.22×; the most common case was skipping 10/20 steps (2.00×). The overall end‑to‑end time still includes text encoding, model preparation, VAE decoding and MP4 packaging, so the speed‑up does not halve total latency.

Cache tiers used:

Draft 0.6 MP – 1056×608 (16:9), 608×1056 (9:16), 768×768 (1:1) – Sage + EasyCache

Standard 0.7 MP – 1152×640 (16:9), 640×1152 (9:16), 832×832 (1:1) – Sage

Fine 0.8 MP – 1216×672 (16:9), 672×1216 (9:16), 896×896 (1:1) – Sage

Standard and fine tiers keep all 20 steps; the native H3 canvas (≈1.0 MP, 1344×768) can be run manually but is not the default for a 24 GB 4090 production line.

Critical Stability Fix: VRAM Debug Barrier

After a 20‑step sampling finishes, the process may stall during the dual VAE decode because the 24 GB VRAM cannot hold all staged sub‑models (text encoder ~14.96 GB, transformer ~20 GB, Video VAE ~5 GB, Audio VAE ~0.58 GB). Insert a VRAM_Debug barrier between sampler and VAE stages with the following settings:

empty_cache       = true
gc_collect        = true
unload_all_models = true

Log example shows freeing ~18 GB, raising free VRAM to ~24 GB before loading the VAE decoders:

free memory before: 5.67 GB
free memory after:  23.97 GB
freed memory:        18.29 GB

Interpreting Performance Data

Recorded end‑to‑end times for 20‑step runs (including VAE) on a RTX 4090:

Resolution 864×480 – Queue→complete 89.10 s – EasyCache skipped 6/20 (1.43×)

Resolution 1056×608 – Queue→complete 120.95 s – EasyCache skipped 7/20 (1.54×)

Resolution 1056×608 – Queue→complete 109.73 s – EasyCache skipped 10/20 (2.00×)

Resolution 608×1056 – Queue→complete 113.53 s – EasyCache skipped 9/20 (1.82×)

Resolution 640×1152 – Queue→complete 120.92 s – EasyCache skipped 10/20 (2.00×)

Resolution 672×1216 – Queue→complete 140.78 s – EasyCache skipped 10/20 (2.00×)

These numbers illustrate the magnitude on a 4090 but are not a strict Sage vs. non‑Sage benchmark. Separate the first run (model loading) from steady‑state production runs and log resolution, frame count, step count, audio presence, cache usage, and whether VAE was re‑loaded.

Common Failure Modes and Direct Remedies

MiniMax H3 node missing: Ensure ComfyUI ≥0.30.0 and both front‑end and back‑end repositories are updated.

Model file present but load error: Verify directory, file size, and SHA256; confirm the .aria2 control file exists to indicate a complete download.

20 steps finish but decode hangs: Place the VRAM unload barrier between sampler and VAE; do not reduce steps prematurely.

Windows desktop stutters: Reserve ~4 GB VRAM for desktop/browser and set ComfyUI process priority to BelowNormal. Run only one GPU worker at a time.

Generated duration slightly longer than expected: Check the 17k+5 frame constraint; 5 s at 24 fps yields 124 frames (≈5.17 s). Adjust post‑processing accordingly.

License Reminder

MiniMax H3 weights are downloadable under a community license that includes regional, commercial, branding, downstream‑service and content‑guarantee clauses (effective from the 2026‑08‑02 release). The license is not equivalent to Apache 2.0; public deployment, paid services, or product integration must respect the official terms.

Final Thoughts

Running H3 on a 24 GB card is stable only when the entire pipeline has clear stage boundaries. Faster nodes alone cannot solve VRAM pressure. Choose a reliable download source with resume and hash verification, keep a single explainable attention/cache strategy, unload large models before decoding, and evaluate benefits using end‑to‑end timings rather than node‑level claims.