Prompt Guide · Updated 2026-07-31

MiniMax Hailuo H3 Prompt Guide

MiniMax H3 — also called Hailuo 3.0, Hailuo 03, or Hailuo-03 — is MiniMax’s multimodal video model: it reads text, images, video, and audio as one input and returns a 5–15 second clip at up to 2K, 24fps, with native stereo audio, generated in a single pass. This guide covers its specs, its omni reference system, and the exact prompt structure it’s built around — available now on Mixio, through the API, or via MCP tool calls.

What is MiniMax H3 (Hailuo 3.0)?

MiniMax H3 is MiniMax’s general-purpose, multimodal video model. Instead of separating generation, editing, and reference-following into different tools, H3 reads text, images, video, and audio as a single context and returns one finished audio-visual clip. It runs in three modes — text-to-video, first-and-last-frame, and omni reference — and is described by MiniMax as open-weight, with weights announced for release.

How is MiniMax H3 different from Hailuo 2.3?

What are MiniMax H3’s full technical specifications?

What is “omni reference” in MiniMax H3?

Omni reference is the mode that changes how you work with H3. One generation accepts up to 9 images, 3 video clips, and 3 audio files — 12 files at most — and each can do a different job: an image sets the character, a second sets the location, a video carries the motion, an audio file carries the voice. Reference video and audio each run 2–15 seconds and 15 seconds in total, and audio can never travel alone — it has to accompany at least one image or video.

Does MiniMax H3 generate audio with the video?

Yes — every H3 generation returns native stereo audio produced in the same pass as the picture. A scene comes back with the line spoken, the footsteps landing, and the room already sounding like a room. Audio is something you direct, not something you accept: a prompt that names instruments, specific effects, and the moment each cue lands returns a materially different clip than one that says nothing about sound.

Can I edit a MiniMax H3 clip without regenerating it?

Yes, via instruction-based editing. Name the one element to change — replace a subject, remove an object, swap a background, relight a scene from day to night, add an effect, or replace a spoken line — and the rest of the frame holds. Several edits can travel in one instruction, which makes an edit pass faster than re-rolling the whole shot and hoping the parts you liked survive.

How does voice cloning and transfer work in MiniMax H3?

A reference recording can give a character a voice they didn’t have, and a supplied line can replace what was said on camera, with the performance adjusted to match. Paired with an identity image, that’s enough to keep one character recognizable across a sequence in both face and sound.

How do I write a MiniMax H3 prompt? (Step-by-step)

H3 rewards a brief that reads like production paperwork, not a caption. Build it in five blocks, in this order:

1. 1

   Assign roles to every reference

   State what each attached image, clip, or audio file is for before describing the action — for example, "Image 1 sets the location, Image 2 is the lead, Audio 1 is the voice." A model given roles does not have to guess which file drives the scene.

2. 2

   Write the beats with timings

   Block the action across the clip's length in seconds — "0–5s she walks the benches, 5–11s she lifts the bottle into the light." Timed beats give MiniMax H3 an order to follow across the full 5–15 second generation instead of averaging the motion.

3. 3

   Fix the look

   Name the style, palette, lighting, and film texture explicitly — lens choice, grain, color temperature, key light direction. This is what separates a shot that reads as filmed from one that reads as generated.

4. 4

   Direct the sound as its own track

   MiniMax H3 generates native stereo audio in the same pass as the picture, so write it like a sound department, not an afterthought — name the instruments, the specific effects, dialogue lines, and the exact moment each cue lands.

5. 5

   State the limits

   Close the brief with what must not change and what must not appear — a locked wardrobe, a face that keeps its hairstyle, no subtitles, no watermark, no modern clothing. Constraints are followed reliably when they are stated, not implied.

What’s the difference between a weak and a strong H3 prompt?

Almost always specificity about time, reference roles, and sound:

What mistakes should I avoid when prompting MiniMax H3?

• Attaching references without saying what each one is for, forcing the model to guess which file drives the scene.
• Describing a frozen frame instead of an action that runs the length of the clip.
• Leaving the audio unwritten, then treating the sound that comes back as a fault of the model.
• Sending an audio reference on its own — it is rejected unless an image or video accompanies it.
• Re-rolling a whole shot to fix one object, when an edit instruction would have kept the rest of the take.

What is MiniMax H3 used for?

How do I generate with MiniMax H3 through the Mixio API or MCP?

One Mixio API key routes to MiniMax H3 and every other supported model — no separate provider account. The same call surface is exposed as MCP tools through the Mixio Gateway, so tool-calling agents like Claude, Cursor, Codex, and Windsurf can drive H3 generations directly from natural language and inherit your production’s locked characters and locations automatically.

curl https://api.mixio.pro/v1/generations \
  -H "Authorization: Bearer $MIXIO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax-h3",
    "mode": "omni-reference",
    "prompt": "0-5s she walks the benches, 5-11s she lifts the bottle into the light",
    "references": [
      { "role": "location", "image_url": "https://cdn.mixio.pro/glasshouse.png" },
      { "role": "talent",   "image_url": "https://cdn.mixio.pro/lead.png" },
      { "role": "voice",    "audio_url": "https://cdn.mixio.pro/vo.wav" }
    ],
    "aspect_ratio": "16:9",
    "resolution": "2k",
    "duration": 12
  }'

See the full interactive model page for live examples and specs: mixio.studio/minimax-h3.

Frequently asked questions