Model · MiniMax · Video-Gen
MiniMax H3
MiniMax's open-weights video model: makes 4–15 second clips with native stereo sound from text, images, video and audio references, with stable dialogue in 11 languages. Strong on instruction following, text and brand rendering, and motion transfer from a reference video.
- Modality
- Video-Gen
- License
- MiniMax H3 Community License Agreement (Open weights)
- Parameter size
- 33B video-and-audio transformer + Qwen3-VL-32B encoder
- Released
- July 31, 2026
- Last verified
- October 28, 2026
- Runs locally
- Yes
- Also handles
- audio-gen
What you need to run it
- Mac
- No. Needs an NVIDIA graphics card — there is no working Mac version.
- Windows / Linux PC
- NVIDIA only. A graphics card with at least 80 GB (320 GB recommended) of video memory. AMD and Intel cards are not supported.
- No graphics card
- Not practical. Too slow without a graphics card or Apple Silicon.
A 33-billion-parameter video model that needs data-centre graphics cards (4 cards in the maker's examples), not an ordinary desktop or Mac.
Numbers are for: 33B model in 16-bit; the maker's sample commands use 4 graphics cards. Our estimate from the model’s size — the maker publishes no figure.
Strengths
- Video with 32 kHz stereo audio in one pass, 4–15 seconds per clip
- Stable spoken dialogue in 11 languages, including English, Chinese, Arabic, Spanish and Japanese
- Omni-reference mode: up to 9 images, 3 video clips and 3 audio clips as references
- Diffusers and SGLang support for self-hosting
Weaknesses
- The licence does not cover the United States, the European Union, the United Kingdom or South Korea — use there needs a separate licence from MiniMax
- Commercial products must prominently show "MiniMax H3" in their interface, and you need written authorisation once those products earn more than $20M a year
- If you let third parties generate with it, you must run abuse safeguards
- Open release outputs 768p by default; the 2K path and the prompt-refinement module are API-only
- Needs data-centre GPUs: the sample serving commands run across 4 GPUs
Try it
| Where | Type | Notes |
|---|---|---|
| Hugging Face | weights | Community licence; excludes the US, EU, UK and South Korea |
| MiniMax API | hosted-api | API key required; includes the 2K workflow |
| Hailuo AI | product | MiniMax's consumer video app |
Official sources
- Announcement announcement
- Model card model
- GitHub github
- Licence docs
Change log
- — Added hardware requirements (Mac support, memory, NVIDIA). Source: an estimate from the 33B model size (66 GB of weights) and the maker's 4-GPU serving examples.
- — Initial entry. Released 2026-07-31 (MiniMax blog) under the MiniMax H3 Community License Agreement, which excludes the US, EU, UK and South Korea, requires a visible "MiniMax H3" credit in commercial products and written authorisation above $20M yearly revenue. Added 2026-10-28 from the popularity sweep: 5.3M Hugging Face downloads in the last month, 6.4k likes, 10.8k GitHub stars.
Esc