MiniMax H3 (local Diffusers)

Open-weight MiniMax H3 via Hugging Face Diffusers ModularPipeline (t2va / fl2va). Runs on your CUDA GPU with ComponentsManager CPU offload by default. Outputs MP4 with native stereo audio.

Hosted API twin: MiniMax H3 (official API) (--model minimax-h3). Fast hosted variant: --model minimax-h3-max (no local twin).

Local weights are H3-Base at 768p (short edge). The hosted H3-Regenerate-2K pass is not open-sourced — use the API for 2K.

License (read this)

The MiniMax H3 Community License for local weights excludes the United States, the European Union, the United Kingdom, and South Korea unless you have a separate authorization. The hosted API (--model minimax-h3) is globally available. Confirm the current terms on the model card before downloading.

Requirements

Item	Notes
GPU	CUDA required. Transformer ~61.7GB bf16 + Qwen3-VL conditioner ~62GB. Offload is on by default.
Host RAM	~75GB+ if you int8-quantize for a 24–32GB consumer card (see Diffusers memory recipes)
Disk	Large snapshot (MiniMaxAI/MiniMax-H3); fl2va loads transformer/, not transformer_ref/
Diffusers	Must install from main — MiniMax-H3 ModularPipeline is not in a stable Diffusers release yet

pip install opentryon[local]
pip install "git+https://github.com/huggingface/diffusers"
export HF_TOKEN=hf_...   # if the repo is gated or rate-limited

Optional overrides:

# export MINIMAX_H3_MODEL_ID=MiniMaxAI/MiniMax-H3
# export MINIMAX_H3_MODEL_PATH=/path/to/local/snapshot

CLI

opentryon video-generate --model minimax-h3-local \
  --prompt "A fashion model walking a runway at dusk, camera tracking, soft ambient sound" \
  --width 960 --height 544 --num-frames 124 --seed 42

opentryon video-generate --model minimax-h3-local \
  --image look.jpg \
  --prompt "Gentle fabric motion as the model turns" \
  --width 960 --height 544 --num-frames 124

width / height must be divisible by 32. Default 960×544 is a practical smaller canvas (~2.3× faster than 1344×768). num_frames snaps to 17*n+5 in the ~5–15s window at 24 fps (124 ≈ 5s).

There is no guidance_scale or negative prompt — the open transformer is CFG-distilled.

Python

from tryon.models import MiniMaxH3LocalAdapter

adapter = MiniMaxH3LocalAdapter()  # MiniMaxAI/MiniMax-H3, CPU offload on
video = adapter.generate_text_to_video(
    prompt="A cinematic shot of fabric flowing in wind, atelier ambience",
    width=960,
    height=544,
    num_frames=124,
    seed=42,
)
open("h3-local.mp4", "wb").write(video)

Notes

• Omni-reference (ref2va / transformer_ref/) is not wired into the CLI; use the hosted API for reference-to-video, or Diffusers directly.
• Consumer-GPU int8 + group offload recipes live in the Diffusers MiniMax-H3 docs.
• MCP tool: video_generate_minimax_h3_local.