🎬 MiniMax-H3 Demo
Omni-modal video + native stereo audio from MiniMaxAI/MiniMax-H3
Showcase official samples · craft H3-style prompts · generate via MiniMax API
Official MiniMax-H3 sample reels
Weights + research card: MiniMaxAI/MiniMax-H3
H3 is a general-purpose omni-modal system: text / image / video / audio in → video with native stereo audio out (4–15s, many aspect ratios, 24 fps, up to 2K via regenerate).
T2VA · text → video+audio (768p)
Text-only prompt → joint video and stereo soundtrack.
T2VA · 2K regenerate
Same story after H3-Regenerate-2K.
FL2VA · first/last frame
Keyframe-conditioned generation with audio.
I2VA · image → video+audio
Single image drives first-frame motion + sound.
I2VA · 2K
Image-conditioned clip regenerated at 2K.
Ref2VA · multi-reference
Images / video / audio references in one packed context.
R2VA · 2K
Reference workflow at higher fidelity.
Direct API · 768p
Open Platform direct 768p reference output.
Direct API · 2K
Open Platform direct 2K reference output.
Local open weights deploy via SGLang / vLLM / diffusers / ComfyUI — see the model card. Unquantized ZeroGPU split demo: multimodalart/minimax-h3.
Build a generation brief the way H3 likes it: subject → action → camera → lighting → environment → audio → style. Dialogue is first-class — H3 supports stable speech in 11+ languages.
Live generation (MiniMax Open Platform)
This tab calls the MiniMax video API (platform-hosted H3 / Hailuo family).
Add secret MINIMAX_API_KEY in the Space settings, or paste a key below (not stored server-side beyond the request).
- API docs: video-generation-v2-create
- Web app: hailuoai.video
MiniMax-H3 at a glance
| Pipeline | image-text-to-video (+ audio) |
| Library | diffusers (MiniMaxH3ModularPipeline) |
| Params | ~33B dense omni-transformer (AdaLN-heavy) |
| Output | 4–15s · 24fps · stereo 32 kHz · multi AR · 768p base · 2K regenerate |
| Variants | FL2VA (text / first-last frame) · Ref2VA (≤9 images, ≤3 videos, ≤3 audios) |
| License | MiniMax H3 Community License |
Local open-weight paths
# diffusers modular pipeline (t2va / fl2va)
pip install -U "diffusers @ git+https://github.com/huggingface/diffusers"
python - <<'PY'
import torch
from diffusers import ModularPipeline
pipe = ModularPipeline.from_pretrained("MiniMaxAI/MiniMax-H3", workflow="t2va")
pipe.load_components(dtype=torch.bfloat16)
print(pipe.doc)
PY
SGLang (multi-GPU) and ComfyUI templates are documented on the model card.
Related Spaces
- multimodalart/minimax-h3 — unquantized ZeroGPU split
- mrfakename/minimax-h3-ultra-fast — NVFP4 ultra path
Built for demoing MiniMaxAI/MiniMax-H3 · not affiliated with MiniMax · samples © MiniMax via the public model repo