Open weights · License
Is MiniMax H3 open source?
No. MiniMax published the H3 weights on Hugging Face under the MiniMax H3 Community License. That makes H3 open weights — downloadable and runnable — but not OSI open source, and not usable under that license in four major markets.
The short answer
- Open weights, custom license: the 33B-parameter H3-Base checkpoints (FL2VA and Ref2VA) are downloadable. The license applies “worldwide, excluding the Excluded Territories”, which it defines as the European Union, the United Kingdom, the Republic of Korea and the United States of America.
- Not the whole system: H3-Context-IR is not included in the release, and H3-Regenerate-2K is not open-sourced. The released weights top out at 768p.
- Commercial terms: products or services above $20 million in yearly revenue need separate prior written authorization from MiniMax, and commercial UIs must prominently display “MiniMax H3”.
- H3 Max and H3 Max Turbo: fal's post-trained variants list no separate weights release. They are API-only, so “open source” does not describe them either.
This page is a factual summary, not legal advice. The license text controls; confirm your own use case before shipping.
Open weights is not the same as open source
The Open Source Definition requires a license that does not discriminate against persons or groups (criterion 5) or against fields of endeavour (criterion 6). A license that excludes four jurisdictions and adds a revenue-triggered authorization step does not meet that standard. MiniMax's model card uses the phrase “open-source release” for the checkpoint drop, but the license it ships with is the part that governs what you may do.
“Open weights” tells you a checkpoint is downloadable. It says nothing about whether you may run it where you live, at your company's size, or inside a competing product. Those are license questions, and they differ sharply across today's open video models.
H3 compared with other open-weight video models
Parameters and license names below come from each official model card. “Not listed” means the card does not claim native audio — it does not mean audio cannot be added in post.
| Model | Developer | Parameters | Weights license | Native audio | Commercial catch |
|---|---|---|---|---|---|
| MiniMax H3 | MiniMax | 33B dense transformer (≈13B AdaLN parameters can be omitted at inference) | MiniMax H3 Community License | Yes — 32 kHz stereo | Not usable under this license in the EU, UK, South Korea or US; over $20M yearly revenue needs written authorization; commercial UIs must display “MiniMax H3”. |
| Wan 2.2 T2V-A14B | Alibaba Tongyi Lab | 27B MoE (14B active per step) | Apache 2.0 | Not listed | No territory or revenue condition in the Apache 2.0 license. |
| HunyuanVideo 1.5 | Tencent | 8.3B | Tencent Hunyuan Community License | Not listed | Does not apply in the EU, UK or South Korea; over 100M monthly active users needs a Tencent license; provider disclosure required. |
| LTX-2 | Lightricks | 19B | LTX-2 Community License | Yes — synchronized video and audio | Commercial use at $10M+ annual revenue needs a paid license; no products that compete with Lightricks without a separate license. |
| Mochi 1 preview | Genmo | 10B | Apache 2.0 | Not listed | No territory or revenue condition in the Apache 2.0 license. |
Parameters and license terms checked on 2026-10-22. See the full source list below.
If you are in an excluded territory
The community license does not authorize use of the downloadable H3 weights in the EU, UK, South Korea or US. MiniMax's license invites parties in those territories to contact it about a separate license, which it says will be granted with controls and guardrails for local law.
A hosted API is a different path from a weight download. This site calls fal's hosted minimax/h3-max-turbo endpoints rather than distributing H3 weights. Hosted-service terms, not the H3 Community License, govern that usage; check fal's and MiniMax's terms for your jurisdiction.
If you need local weights with no territory clause, the Apache 2.0 options in the table — Wan 2.2 and Mochi 1 — are the lowest-friction starting points. Each has its own capability trade-offs.
Self-host or use the hosted generator?
H3 is a 33B dense transformer, and MiniMax's own serving example uses multiple GPUs for a single checkpoint. The full breakdown of what is and is not in the release is on the H3 open-weights page.
If you need clips rather than a GPU fleet, this site runs H3 Max Turbo through fal's API. You sign in, add credits, and generate 5–15 second videos at 480P or 768P with native audio. Generation requires purchased credits; there are no free signup credits.
MiniMax H3 open-source FAQ
Is MiniMax H3 open source?
No. MiniMax released the H3 weights under the MiniMax H3 Community License, a custom license with territorial and commercial restrictions. It is open weights, not an OSI-approved open-source license.
Can I use MiniMax H3 in the United States, EU, UK or South Korea?
Not under the community license. The license defines the Applicable Territory as worldwide, excluding the European Union, the United Kingdom, the Republic of Korea and the United States of America. MiniMax says parties in those territories can contact it about a separate license.
What is the MiniMax H3 revenue threshold?
Commercial products and services that generate more than $20 million in yearly revenue need separate prior written authorization from MiniMax. Below that threshold the community license still applies, including its attribution requirement.
Do I have to credit MiniMax H3 in my product?
Yes. The license requires commercial products or services that use MiniMax H3 to prominently display “MiniMax H3” in the user interface.
Can I fine-tune or distill MiniMax H3?
You may create Model Derivatives within the Applicable Territory, but the license says you may not use H3 or its outputs to improve any other AI model other than H3 or its Model Derivatives. Review the license text before training on H3 outputs.
Are H3 Max and H3 Max Turbo open source?
No. H3 Max and H3 Max Turbo are fal's post-trained variants of H3. Their model pages list no separate weights release; they are served through fal's API. The open weights cover H3-Base, in its FL2VA and Ref2VA variants.
Which open-weight video models have the most permissive license?
Wan 2.2 T2V-A14B and Mochi 1 preview are published under Apache 2.0. MiniMax H3, HunyuanVideo 1.5 and LTX-2 each use a custom community license with territory, revenue or competing-product conditions.
What is missing from the open MiniMax H3 release?
H3-Context-IR, the hosted multimodal preprocessing system, is not included, and H3-Regenerate-2K is not open-sourced, so the released weights generate at 768p rather than 2K. The initial release also uses full attention; MiniMax says sparse attention will follow in a later update.
How can I use H3 Max Turbo without a GPU?
Use a hosted service. This site runs H3 Max Turbo through fal's API: you sign in, add credits, and generate 5–15 second clips at 480P or 768P with native audio. Generation requires purchased credits; no free signup credits are granted.
Sources (accessed 2026-10-22)
- MiniMax AI — MiniMax H3 model card (Hugging Face)
- MiniMax AI — MiniMax H3 Community License (full text) — Excluded Territories in clause 1.5; commercial terms in section IV
- Open Source Initiative — The Open Source Definition — Criteria 5 and 6: no discrimination against persons, groups or fields of endeavour
- fal — MiniMax H3 vs. MiniMax H3 Max — Published September 1, 2026
- Alibaba Tongyi Lab — Wan 2.2 T2V-A14B model card
- Tencent — HunyuanVideo 1.5 model card
- Tencent — Hunyuan Community License (full text)
- Lightricks — LTX-2 model card
- Lightricks — LTX-2 Community License (full text)
- Genmo — Mochi 1 preview model card