MiniMax H3 (official API)

First-party MiniMax H3 (also marketed as Hailuo 3 / Hailuo-03) video generation with native stereo audio. Uses the V2 endpoints (POST /v2/video_generation, GET /v2/query/video_generation/{task_id}). This is a different surface from Hailuo 2.3, which stays on V1.

Switch H3 vs H3 Max (fast variant) with the CLI model id. Same adapter class, same MINIMAX_API_KEY. H3 Max has no open-weight twin.

CLI model	API model	Modes	Resolution	Duration
minimax-h3	MiniMax-H3	T2V, first/last-frame I2V, reference-to-video (Python)	768P, 2K (default 2K)	4–15s
minimax-h3-max	MiniMax-H3-Max	T2V, first/last-frame I2V only (no R2V)	480P, 768P (default 768P; no 2K)	5–15s

Local open weights: MiniMax H3 (local Diffusers) (--model minimax-h3-local). Local is 768p H3-Base only; this API path is the hosted 2K workflow. H3 Max is API-only.

H3 on the API is billed as pay-as-you-go video.

Environment

export MINIMAX_API_KEY=...   # same key as hailuo-2.3
# export MINIMAX_API_BASE_URL=https://api.minimax.io

CLI

opentryon video-generate --model minimax-h3 \
  --prompt "A fashion model walking a runway at dusk, camera tracking" \
  --duration 5 --resolution 2K --ratio 16:9

opentryon video-generate --model minimax-h3 \
  --image look.jpg --prompt "Gentle fabric motion as the model turns" \
  --duration 6 --resolution 2K

# Last-frame only, or first + last
opentryon video-generate --model minimax-h3 \
  --prompt "Walk toward camera, stop on the mark" \
  --last-frame end.jpg --duration 5

# Fast variant (H3 Max) — 768P default, no 2K / no reference-to-video
opentryon video-generate --model minimax-h3-max \
  --prompt "A fashion model walking a runway at dusk, camera tracking" \
  --duration 5 --resolution 768P --ratio 16:9

Flag	H3 (minimax-h3)	H3 Max (minimax-h3-max)
--duration	Integer 4–15 seconds (default 5)	Integer 5–15 seconds (default 5; 4s invalid)
--resolution	768P or 2K (default 2K)	480P or 768P (default 768P)
--ratio	Required for T2V: 21:9, 16:9, 4:3, 1:1, 3:4, 9:16. Cannot be adaptive. I2V ignores ratio (always adaptive from the keyframe).	Same
--image	First frame (switches to I2V)	Same
--last-frame	Last frame (alone or with --image)	Same

Prompt is required (max 7000 characters).

Python

from tryon.api.minimax import MiniMaxH3Adapter

adapter = MiniMaxH3Adapter()
video = adapter.generate_text_to_video(
    prompt="A fashion model walking a runway at dusk, camera tracking",
    duration=5,
    resolution="2K",
    ratio="16:9",
)
open("h3.mp4", "wb").write(video)

# First + last frame
video = adapter.generate_image_to_video(
    image="look.jpg",
    prompt="Gentle turn toward the camera",
    last_frame="end.jpg",
    duration=6,
)

# Reference-to-video (H3 only; mutually exclusive with first/last frames)
video = adapter.generate_text_to_video(
    prompt="Character speaks: follow the wind. Voice follows reference audio 1.",
    reference_image=["subject.jpg"],
    reference_audio=["voice.mp3"],
    duration=5,
    resolution="2K",
)

# Fast variant
fast = MiniMaxH3Adapter(model="MiniMax-H3-Max")
open("h3-max.mp4", "wb").write(
    fast.generate_text_to_video(
        prompt="Runway walk at dusk",
        duration=5,
        resolution="768P",
    )
)

Notes

• Auth is Authorization: Bearer {MINIMAX_API_KEY}.
• Jobs are async: create returns task_id, poll until succeeded, then download task.content.url (time-limited).
• Image-to-video and reference-to-video cannot be mixed in one request.
• H3 Max on MiniMax does not support reference image / video / audio — use --model minimax-h3, or Fal --model fal-h3-max (Fal H3 Max).
• MCP tools: video_generate_minimax_h3, video_generate_minimax_h3_max (generated from the registry — no hand-written wrappers).