Back to Gap Map Model components / Image, video, 3D & music generation

MiniMax Hailuo

open weights / Overall score: 4.6(leading)

MiniMax Hailuo is MiniMax's video generation line. Its downloadable model, MiniMax H3, is a 33B single-stream transformer that generates 4 to 15 second clips with 32 kHz stereo audio from text, first and last frames, or up to nine reference images, three video clips and three audio clips. The Hailuo 02 and 2.3 models, and the H3 2K regeneration and context modules, are served only through MiniMax's app and API.

The Hugging Face download count for MiniMax H3 is unusually high for a model released weeks earlier; it was re-read and is recorded as published.

Openness

3 high confidence

3.0

weights: open(MiniMax-H3 downloads from Hugging Face without a gate
data: closed(the README names no training data)
code: partial(inference code only, with full attention
license: MiniMax-H3-Community-License(licensed for a territory excluding the European Union, the United Kingdom, the Republic of Korea and the United States

MiniMax's community license permits commercial use but excludes the European Union, the United Kingdom, South Korea and the United States entirely, and requires written permission above USD 20M in yearly revenue. Only inference code is published and the training data is not described. The MiniMax H3 Community License bounds commercial use by revenue and excludes several territories, and a territorial exclusion caps a license at use_bounded.

https://huggingface.co/api/models/MiniMaxAI/MiniMax-H3?expand[]=downloads&expand[]=likes&expand[]=cardData&expand[]=gated&expand[]=lastModified&expand[]=createdAt recorded 2026-09-26
Hugging Face API record: gated false, license minimax-h3-community-license-agreement.
https://huggingface.co/MiniMaxAI/MiniMax-H3/raw/main/LICENSE recorded 2026-09-26
License body: "MiniMax H3 COMMUNITY LICENSE AGREEMENT"; "Excluded Territories" means the European Union, the United Kingdom, the Republic of Korea and the United States of America; a separate authorization is needed above "20 million US dollars" in yearly revenue.
https://raw.githubusercontent.com/MiniMax-AI/MiniMax-H3/main/README.md recorded 2026-09-26
README: "The initial open-source release provides inference with full attention only"; H3-Context-IR "is not included in this open-source release"; no training data is named.

Adoption

4 high confidence

4.0

Adoption is measured as Hugging Face downloads of the single MiniMax H3 repository, which holds both of its checkpoints.

https://huggingface.co/api/models/MiniMaxAI/MiniMax-H3?expand[]=downloads&expand[]=likes&expand[]=cardData&expand[]=gated&expand[]=lastModified&expand[]=createdAt recorded 2026-09-26
Hugging Face API record for MiniMaxAI/MiniMax-H3: downloads 3657004, createdAt 2026-07-28.

Capability

5 medium confidence

5.0

MiniMax H3 is among the three strongest video models on the arena, closed models included, and the strongest with downloadable weights. Only two hosted models place above it.

https://artificialanalysis.ai/video/leaderboard/text-to-video recorded 2026-09-26
Arena FAQ: "The top Text to Video models without audio by Elo rating are: 1. Wan 3.0 (Elo 1335), 2. Gemini Omni Flash (Elo 1332), 3. MiniMax H3 (Elo 1302)".

Verified 2026-09-26