MiniMax H3

Omni-modal video generation with native stereo audio and 2K output

About model

MiniMax H3 is a general-purpose omni-modal generation model, the third generation of MiniMax's video line following Hailuo 01 and 02. It understands unified context spanning text, images, video, and audio, and generates 4 to 15 second clips with native stereo sound at up to 2K resolution. Where earlier generative systems split work across separate expert models for text-to-video, image-to-video, first-and-last-frame, reference, and editing, H3 unifies these as one model, with reference and editing relationships expressed in natural language: describe how the inputs relate to the target video, and H3 handles the cross-modal understanding itself. MiniMax reports early testing shows the model ready for commercial content creation, with strong instruction following, accurate text and brand rendering, and video-to-video motion transfer. Weights are open under the MiniMax H3 Community License. Available on Together AI.

‍

Max Resolution

In-context regeneration recovers small text and fine detail that super-resolution can only guess at

Native Audio

Stereo

Voice, sound effects, and music jointly generated at 32kHz, with stable dialogue in 11 languages

Max Reference Files

Up to 9 images, 3 video clips, and 3 audio clips combined in a single request

Model key capabilities

Omni-Modal Reference & Editing: Blends up to 9 images, 3 videos, and 3 audio clips as inputs, with relationships described in natural language rather than fixed task types
Native Stereo Audio: Voice, sound effects, and music jointly modeled and generated with the video, with stable dialogue support across 11 languages
2K In-Context Regeneration: Regenerates its own 768p output at 2K using the original context, recovering small text and fine detail
Available on Together AI: Serverless access to the full generation pipeline through the Together AI API

API usage

cURL
Python
Typescript

Endpoint:

MiniMaxAI/MiniMax-H3

curl --request POST \
  --url https://api.together.xyz/v2/videos \
  --header "Authorization: Bearer $TOGETHER_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "model": "MiniMaxAI/MiniMax-H3",
    "prompt": "some penguins building a snowman"
  }'

from together import Together

client = Together()

# Create a video generation job
job = client.videos.create(
    prompt="A serene sunset over the ocean with gentle waves",
    model="MiniMaxAI/MiniMax-H3"
)

import Together from "together-ai";

const together = new Together();

async function main() {
  // Create a video generation job
  const job = await together.videos.create({
    prompt: "A serene sunset over the ocean with gentle waves",
    model: "MiniMaxAI/MiniMax-H3"
  });

Model card
Architecture Overview:
• H3-Omni Transformer: a 33B parameter dense, single-stream Transformer with no modality-specific attention or FFN structures; roughly 13B parameters sit in AdaLN branches whose outputs can be precomputed for inference
• Multimodal encoding built on a vision-language encoder derived from Qwen3-VL-32B, alongside H3-VisualVAE (16x spatial, 4x temporal compression, 24 latent channels) and a stereo H3-AudioVAE compressing 32kHz audio to 40Hz latents
• Three-dimensional Multimodal RoPE captures temporal and spatial relationships across the packed multimodal sequence
• A hosted context-understanding stage interprets free-form multimodal instructions, parsing cross-modal and temporal relationships before generation
• 2K output produced by in-context regeneration: the base model regenerates its own 768p result while re-reading the original context, rather than using a separate super-resolution module

Training Methodology:
• Unified pre-training across text-to-image, text-to-video, text-to-audio, and generalized reference and editing tasks, built from natural data with language as the generalization bridge
• A purpose-built full-modality captioning pipeline distills roughly 100K tokens of inference per source item into ~4K-token contextual descriptions covering relationships among context elements
• Training separates understanding and generation workloads to balance heterogeneous compute, with a MiniMax-reported throughput gain of nearly 30%

Performance Characteristics:
• MiniMax reports early testing shows commercial readiness across advertising, branding, e-commerce, product design, UI/UX, and gaming content
• Strengths highlighted by MiniMax include instruction following, accurate text and brand rendering, and video-to-video motion transfer
• Output specifications: 4-15 second clips at 24 FPS, aspect ratios from 21:9 to 9:16, 768p default with 2K regeneration, 32kHz stereo audio
‍
Prompting
Together AI API Access:
• Access MiniMax H3 via Together AI APIs using the endpoint MiniMaxAI/MiniMax-H3
• Authenticate using your Together AI API key in request headers
• Create a video request with a text prompt, optionally attaching images, video clips, and audio references, then poll until the video completes
• Describe reference and editing relationships in natural language, such as referencing camera movement from one video while matching vocals to a supplied audio clip
• Outputs run 4 to 15 seconds with native stereo audio, at 768p by default with 2K available
• Available on Together AI with serverless access
‍
Applications & use cases
Advertising & E-Commerce:
• Produce product and campaign video with accurate on-screen text and brand rendering
• Keep products and brand elements consistent using multi-image reference inputs
• Generate localized variants with native dialogue across 11 languages

Film & Motion Design:
• Create opening titles, animated posters, and motion pieces with synchronized sound design
• Transfer camera movement and motion from reference footage to new generations
• Edit existing clips through natural language, from re-voicing characters to changing scenes

Product & Interface Showcase:
• Turn product designs and UI mockups into polished demonstration video
• Animate websites, apps, and game concepts with matched music and effects
• Iterate on creative directions quickly from the same reference set
‍

Related models

Model specifications

Model data

Model provider
MiniMax AI
Type
Video
Resolution/Duration
Up to 2K, 4-15s
Deployment
Serverless
Endpoint
MiniMaxAI/MiniMax-H3
Parameters
33B
Price
$0.1391 / video
Input modalities
Text
Image
Video
Audio
Output modalities
Video

Released
July 30, 2026
External link
Provider docs
Category
Video

Run in Playground

Quickstart docs

Deploy model

MiniMax H3

About model

API usage

Model card

Prompting

Applications & use cases