MiniMax H3 (Hailuo-3): Open-Weights Multimodal Video AI
Generates 2K native videos up to 15s with synchronized stereo audio, V2V motion transfer, and unified cross-modal editing.
Text and audio AI progressed fast, but AI video remained stale and trapped in closed-source silos.
MiniMax H3 (internally named Hailuo-3, where Hailuo 海螺 means "sea snail" in Chinese) disrupts this space:
- 🔓 Open-Weights Ecosystem: Self-hostable model weights releasing soon on Hugging Face.
- 💰 Disruptive Pricing: Starting from $0.13 / second (via OpenRouter & MiniMax API).
- 🎙️ Native Stereo Sound: Generates dual-channel audio matched directly to visual scene acoustics.
- Visit platform.minimax.io.
- Log in and open Console.
- Go to Access ➔ API Key.
- Copy your key and paste it into
.env:
MINIMAX_API_KEY=your_api_key_here# 1. Environment Setup
python -m venv venv
.\venv\Scripts\Activate.ps1
pip install -r requirements.txt
# 2. Configure Key & Run Task Generator
Set-Content -Path .env -Value "MINIMAX_API_KEY=your_api_key_here"
python generate_video.py┌────────────────────────────────────────────────────────────────────────┐
│ MiniMax H3 Model Card │
├───────────────────┬────────────────────────────────────────────────────┤
│ Output Resolution │ Up to 2K (Native audiovisual generation) │
│ Duration │ 4–15 seconds (Integer seconds per request) │
│ Aspect Ratios │ 21:9 · 16:9 · 4:3 · 1:1 · 3:4 · 9:16 (Adaptive) │
│ Multimodal Input │ Up to 9 images, 3 videos, 3 audio clips (≤12 total)│
│ Core Strengths │ Instruction edits, text/brand rendering, V2V │
│ Model Weights │ Open (Self-hostable on H100s + hosted API) │
│ Commercial Pricing│ From $0.13 / second │
└───────────────────┴────────────────────────────────────────────────────┘
Community hands-on testing reveals standout capabilities:
- ⏱️ Generation Latency: ~2 minutes for Text-to-Video; ~5 minutes for Image-to-Video.
- 👤 Identity Preservation: Holds facial features, hair, and apparel consistently across frames without drift.
- 🖐️ Anatomy Fidelity: Renders natural hands, facial expressions, and water reflections accurately.
- 🚗 Reference-to-Video: Re-synthesizes camera trajectories and racing scenes with realistic audio stems.
MiniMax H3 unifies text, image, video, and audio tasks into a single pretraining framework.
┌────────────────────────────────────────────────────────┐
│ UNIFIED MULTIMODAL INPUT │
│ Text Prompt + Images + Video + Audio Stems │
└───────────────────────────┬────────────────────────────┘
│
▼
┌────────────────────────────────────────────────────────┐
│ 1. Contextual Omni Representation │
│ (100K token inference ➔ 4K distilled) │
└───────────────────────────┬────────────────────────────┘
│
▼
┌────────────────────────────────────────────────────────┐
│ 2. H3-Omni Transformer │
│ (Separated Understanding & Generation Compute) │
└───────────────────────────┬────────────────────────────┘
│
▼
┌────────────────────────────────────────────────────────┐
│ 3. H3-VAE & In-Context Regen │
│ (4x Sequence Compression ➔ 2K Native Video) │
└───────────────────────────┬────────────────────────────┘
| Core Technology | Technical Impact & Mechanism |
|---|---|
| Contextual Omni Rep | Unifies cross-modal relationships; distills 100K inference tokens to 4K. |
| H3-VAE Tokenizer | 4x sequence compression ratio powers native 2K resolution efficiently. |
| H3-Omni Transformer | Separates understanding and generation workloads, boosting training speed by ~30%. |
| In-Context Regen | Base model regenerates low-res frames in-context at 2K to restore fine text & detail. |
| Mode | Input Assets | Typical Commercial Application |
|---|---|---|
| Text-to-Video | Text Prompt | Generate 2K video scenes from scratch |
| First/Last-Frame I2V | Prompt + Start/End Image | Control exact starting and ending video keyframes |
| Reference Generation | Prompt + Images/Videos/Audio | Match subject identity, camera motion, style, or voice |
| Item | Requirement |
|---|---|
| Reference Images | Up to 9 images (Width/Height: [256, 5760]; ≤ 30 MB per file) |
| Reference Videos | Up to 3 clips (Duration: 2–15s; ≤ 50 MB per file) |
| Audio Clips | Up to 3 clips (Duration: 2–15s; ≤ 15 MB per file) |
| Combined Cap | Max 12 files total (API request body ≤ 64 MB) |
| Prompt Limit | ≤ 7,000 characters |
[ POST /v2/video_generation ] ──► [ GET /v2/video_generation/{task_id} ] ──► [ Download MP4 ]
- Submit Task: POST payload with prompt and resolution parameters to receive
task_id. - Poll Status: GET task status until state equals
Success. - Fetch Video: Download MP4 video from returned
content.url.
.env: StoresMINIMAX_API_KEY.generate_video.py: Submits generation tasks to MiniMax H3 REST API.README.md: Visual project documentation.requirements.txt: Dependencies (requests,python-dotenv).
- Film Opening Titles: Render cinematic 2K scenes with camera tracking and sound effects.
- E-Commerce Showcases: Generate 3D product animations with sharp brand rendering.
- Social Video Campaigns: Produce 16:9 and 9:16 marketing videos with voice stems.
- V2V Motion Transfer: Transfer video choreography onto custom visual characters.
- UI/UX Motion Demos: Animate mobile app micro-interactions and glassmorphic interfaces.
- Auto-Downloader: Loop status polling to automatically save
output.mp4. - Reference Asset CLI: Pass image, video, and audio reference paths in CLI arguments.
- Multi-Shot Composer: Stitch 15s clips into full-length commercial videos.
- Hugging Face Integration: Local inference support once open weights drop.
- Batch Exporter: Parallel 2K and 768p multi-resolution rendering.
MiniMax H3 Hailuo 3 MiniMax AI Open Source Video AI 2K Video Generation Text to Video AI Image to Video Multimodal AI Stereo Sound Generation V2V Motion Transfer AI Video Model MiniMax API
MIT License. Built for AI creators and developers leveraging MiniMax H3.