By MiniMaxLearning Objectives
- Explain what makes H3 different from video models that generate silent clips
- Determine whether you are actually licensed to download and run the weights
- Compare the open-weight and hosted-API paths for generative video
What Is MiniMax H3?
MiniMax H3 is an omni-modal generative video model from the Shanghai lab MiniMax, launched on July 31, 2026, with weights published to Hugging Face on August 2. It reads text, images, video and audio in a single context and returns video with matching audio generated at the same time — not a silent clip you score afterwards. That is the substance of the release: most downloadable video models produce picture only.
The system is a 33-billion-parameter dense Transformer built from three staged components — H3-Context-IR for preprocessing, H3-Base for generation, and H3-Regenerate-2K for upscaling — and ships in two task-specific variants, FL2VA (first-and-last-frame) and Ref2VA (reference-to-video).
⚠️Warning
Read this before you plan a deployment: the weights are not licensed for use in the United States. The MiniMax H3 Community License defines "Excluded Territories" as the European Union, the United Kingdom, the Republic of Korea and the United States of America, and states that you "may not use, reproduce, modify, distribute, or display the MiniMax H3 Works or any of their Outputs or results outside the Applicable Territory." MiniMax attributes the carve-out to the regulatory picture for generative video — the EU AI Act, copyright uncertainty in the UK and South Korea, and active US litigation. Organizations in the excluded regions can apply to MiniMax for a formal license, which the company reviews case by case. The hosted API is globally available and is unaffected by this restriction.
A good deal of launch coverage described H3 simply as an open-weight release with a $20 million revenue cap. That is incomplete in the way that matters most to a US reader.
Capabilities
| Property | Value |
|---|---|
| Parameters | 33 billion (dense Transformer) |
| Clip length | 4 to 15 seconds |
| Resolution | Up to 2K; default shorter side 768 pixels |
| Frame rate | 24 frames per second |
| Audio | 32 kilohertz stereo, generated jointly with the video |
| Aspect ratios | 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 |
| Inputs | Text, images, video and audio in one context |
| Modes | Text-to-video, first-and-last-frame, reference-to-video, video editing |
MiniMax positions H3 strongest at video editing, where it claims the leading result. On generation it is more modest: independent comparisons place it behind Google's Gemini Omni Flash on text-to-video, and behind both Gemini Omni Flash and ByteDance's Seedance on image-to-video. Joint audio, not raw generation quality, is the reason to reach for it.
💡Key Concept
Why joint audio-video matters. Generating sound separately from picture means lip movements, footsteps and object collisions have to be re-synchronized by hand, which is most of the post-production work on a short clip. A model that produces both in one pass from a shared representation gets that alignment approximately right by construction — which is worth more to a production workflow than a couple of points of visual fidelity.
Licensing
| Term | What it says |
|---|---|
| Excluded Territories | European Union, United Kingdom, South Korea, United States — no use of weights or their outputs |
| Revenue threshold | Above $20 million yearly revenue, you need prior written authorization from MiniMax |
| Attribution | Must prominently display 'MiniMax H3' in the user interface of a commercial product |
| Model training | Outputs may not be used to improve any other AI model, except H3 itself or its derivatives |
| Hosted API | Globally available, not subject to the territory restriction |
This is not an open-source license, and the catalog does not classify it as one. It is a source-available community license with a geographic carve-out that happens to exclude most of this platform's readership from the self-hosting path.
Pricing
- Excluded in the US, EU, UK and South Korea
- Attribution required in commercial UI
- Written authorization above $20 million revenue
- Globally available including the US
- No territory restriction
- MiniMax-operated safeguards
- For organizations in excluded territories
- Case-by-case review by MiniMax
- Requires documented compliance controls
Strengths
- Joint audio and video — 32 kilohertz stereo generated with the picture, which very few downloadable models do at all
- Genuinely published weights — a 499 GB release in diffusers format, not a promise
- Strong at editing — MiniMax's best claimed result, and the most practical mode for existing footage
- Wide input flexibility — text, image, video and audio in one context, with first-and-last-frame and reference-to-video modes
- Full aspect-ratio range — 21:9 through 9:16 covers cinematic and vertical-social in one model
- API unaffected by the territory limits — the hosted path is open to everyone
Limitations & Considerations
- Not licensed for weights use in the US, EU, UK or South Korea — the single most important constraint for most readers here
- Not open source, despite widespread description as such — attribution, revenue and territory conditions all apply
- Behind on pure generation — trails Gemini Omni Flash on text-to-video, and Gemini Omni Flash and Seedance on image-to-video
- Short clips — 4 to 15 seconds, against the 30-second clips Seedance 2.5 ships
- Heavy to self-host — a 499 GB multi-component release is a serious infrastructure commitment
- No-compete clause — outputs cannot be used to train or improve any other model
- Data residency — the hosted API runs on MiniMax infrastructure in China
Best Use Cases
| Task | Why H3 |
|---|---|
| Short clips that need sound | Joint 32 kilohertz stereo generation avoids manual re-synchronization |
| Editing existing footage | MiniMax's strongest claimed capability |
| Reference-driven generation | The Ref2VA variant is purpose-built for it |
| Self-hosted video generation, outside excluded territories | One of very few genuinely downloadable audio-and-video models |
| Vertical social formats | Native 9:16 and 3:4 without cropping |
When to choose alternatives:
- Highest text-to-video quality → Google Veo 3 or Gemini Omni
- Longer clips → ByteDance Seedance 2.5, at 30 seconds
- US-based self-hosting → an actually permissive model; H3's license does not allow it
- Commercial work needing clean license terms → Veo, Runway or another vendor with straightforward commercial rights
Getting Started
- Check your territory first. If you are in the US, EU, UK or South Korea, the weights are not available to you — use the API, or apply to MiniMax for a formal license
- Try the hosted model on the Hailuo platform to judge quality before any infrastructure work
- If you are licensed for the weights, review the repo at huggingface.co/MiniMaxAI/MiniMax-H3 and budget for a 499 GB multi-component download
- Start with the editing modes rather than text-to-video — that is where H3 is strongest relative to rivals
- If you ship commercially, wire the required "MiniMax H3" attribution into your interface from the start
Key Takeaways
- MiniMax H3 is a 33-billion-parameter video model that generates picture and synchronized 32 kilohertz stereo audio together, with weights genuinely published on Hugging Face
- Its license excludes the United States, European Union, United Kingdom and South Korea from using the weights at all — the hosted API remains globally available
- It is not open source: attribution is mandatory, revenue above $20 million requires written authorization, and outputs may not be used to improve other models
- It leads on video editing but trails Gemini Omni Flash and Seedance on text-to-video and image-to-video generation
- Clips run 4 to 15 seconds at up to 2K and 24 frames per second, across aspect ratios from 21:9 to 9:16
