Run and self-host MiniMax H3

​

Status and scope

​

Choose a deployment path

​

Run H3 locally with ComfyUI

​

ComfyUI baseline

​

Quickstart in the Template Library

1. Update ComfyUI to version 0.30.0 or later and start it normally.
2. Open Template Library > Video.
3. Select MiniMax H3 T2V, MiniMax H3 I2V, or MiniMax H3 R2V.
4. Follow the model-scan pop-up to download the required files. Restart ComfyUI if the model list does not refresh.
5. Set the prompt and any input image, video, or audio references.
6. In Resolution Selector, use 0.98 megapixels with a multiple of 32, or set 1344 × 768 directly for a 16:9 native canvas. Do not use the 1.0 megapixel step: it resolves to 1376 × 768, above the H3-Base 768 × 1344 pixel-area limit.
7. Queue the workflow. The result should be an MP4 containing a 24 FPS video stream and synchronized stereo audio.

​

Choose a workflow

​

Manual model placement

​

Resolution, duration, and Turbo mode

• H3-Base uses a 768-pixel short edge and dimensions aligned to a multiple of 32; 1344 × 768 is the native 16:9 canvas.
• Duration snaps to the model’s 17k + 5 frame grid at 24 FPS. A requested duration can therefore be adjusted to a valid frame count.
• The example FL2VA workflow uses 20 steps by default. Its optional 8-step Turbo LoRA is faster but can reduce motion and audio quality.
• The R2V workflow uses a separate optional 4-step Turbo LoRA. Record the LoRA name, strength, steps, seed, and all references when comparing results.

​

SGLang hardware and environment

​

SGLang Quickstart

​

1. Install the pinned SGLang source

export MINIMAX_H3_WORKDIR="$PWD/minimax-h3-deploy"
mkdir -p "$MINIMAX_H3_WORKDIR"

git clone https://github.com/sgl-project/sglang.git "$MINIMAX_H3_WORKDIR/sglang"
cd "$MINIMAX_H3_WORKDIR/sglang"
git checkout 2511743bd784e69e5a81ca3d926a000711dae4ab

python3 -m pip install --upgrade uv
uv venv --python 3.12 .venv
source .venv/bin/activate
uv pip install -e "python[diffusion]" --prerelease=allow
uv pip install "huggingface_hub==1.28.0"
sglang --version

​

2. Download the pinned FL2VA checkpoint

export HF_HOME="$MINIMAX_H3_WORKDIR/hf-cache"
export MINIMAX_H3_MODEL_DIR="$MINIMAX_H3_WORKDIR/model"

hf download MiniMaxAI/MiniMax-H3 \
  --revision 42ed227ee7df40d41602854ae760620d6eb651fe \
  --include "model_index.json" "FL2VA/*" \
  --local-dir "$MINIMAX_H3_MODEL_DIR"

​

3. Start the FL2VA service

HF_HUB_OFFLINE=1 sglang serve \
  --model-path "$MINIMAX_H3_MODEL_DIR" \
  --model-variant fl2va \
  --num-gpus 8 \
  --tp-size 1 \
  --ulysses-degree 8 \
  --encoder-parallel auto \
  --performance-mode speed \
  --host 127.0.0.1 \
  --port 30010

​

4. Check readiness

curl --fail --retry 60 --retry-delay 10 --retry-all-errors \
  http://127.0.0.1:30010/health

{"status":"ok"}

​

5. Generate, poll, and download an MP4

video_id=$(
  curl -fsS -X POST http://127.0.0.1:30010/v1/videos \
    -H "Content-Type: application/json" \
    -d '{
      "model": "MiniMaxAI/MiniMax-H3",
      "prompt": "At night, three cats perform with tiny brass instruments in a warm living room, with movement synchronized to the music.",
      "seconds": 5,
      "task": "t2va",
      "conditions": [],
      "target": {
        "short_edge": 768,
        "aspect_ratio": "16:9",
        "duration_seconds": 5.0
      },
      "num_outputs_per_prompt": 1,
      "num_inference_steps": 50,
      "flow_shift": 12.0,
      "audio_flow_shift": 3.0,
      "seed": 1101
    }' | jq -r '.id'
)

test -n "$video_id" && test "$video_id" != "null"

while true; do
  job=$(curl -fsS "http://127.0.0.1:30010/v1/videos/${video_id}")
  status=$(printf '%s' "$job" | jq -r '.status')

  case "$status" in
    completed) break ;;
    failed)
      printf '%s\n' "$job" | jq . >&2
      exit 1
      ;;
    *) sleep 5 ;;
  esac
done

curl -fsS -L \
  "http://127.0.0.1:30010/v1/videos/${video_id}/content" \
  -o minimax-h3-t2va.mp4

ffprobe -v error \
  -show_entries stream=codec_type,codec_name,width,height,r_frame_rate,sample_rate,channels \
  -of json minimax-h3-t2va.mp4

​

Capability and checkpoint matrix

​

SGLang hardware and performance data

​

SGLang weights, cache, and offline deployment

• The official Hugging Face repository is MiniMaxAI/MiniMax-H3. The default cache root is ~/.cache/huggingface; this page overrides it with HF_HOME.
• The mainland China mirror is MiniMax/MiniMax-H3. Revision parity with the pinned Hugging Face commit has not been documented, so it is not used for the reproducible baseline.
• For an air-gapped host, run the pinned hf download command on a connected machine, transfer the complete model directory, and verify that model_index.json plus the selected FL2VA/ or Ref2VA/ directory are present. Start SGLang with HF_HUB_OFFLINE=1.
• Interrupted Hugging Face downloads are resumable. Re-run the same command with the same revision. If a cached file is corrupt, use hf download --force-download for that pinned revision rather than switching revisions silently.
• Do not download both checkpoint partitions unless the deployment serves both; each partition is an independent service.

​

SGLang production and security

• Put it behind an authenticated reverse proxy or API gateway; require an API key or equivalent identity control.
• Terminate TLS at the ingress and restrict the backend port with a firewall, security group, or private network.
• Apply request quotas, concurrent-job limits, output retention, audit logging, and abuse controls required by the H3 license.
• Mount server-local reference media read-only and allow only dedicated directories. Do not expose arbitrary filesystem paths.
• If an ingress downloads remote image, video, or audio URLs, enforce a domain allowlist, redirect limit, connection/read timeout, maximum bytes, decoded image dimensions, and media duration before passing a local file to SGLang.

​

SGLang troubleshooting

​

License and system boundary