# MiniMax (Hailuo)

> MiniMax is the Shanghai AI company behind Hailuo video, Talkie and the M-series models. Its H3 video model, open weights, license and Hong Kong listing.

Source: https://metavert.io/minimax  
Published: 2026-10-21  
Updated: 2026-10-21

**MiniMax** is a Shanghai AI company that builds language, speech, music and video models and the consumer apps around them. Its products include the Hailuo AI video platform and the AI character apps Talkie and Xingye. It was founded in December 2021 by former SenseTime computer-vision researchers, with Yan Junjie as CEO. Alibaba led a $600 million round in March 2024 that valued it at $2.5 billion; Tencent is also an investor. On January 9, 2026 MiniMax listed on the Hong Kong Stock Exchange (0100.HK), raising about $619 million at HK$165 a share, the top of its range, for a valuation of roughly $6.5 billion. Its shares more than doubled on the first day. In video it's one of the Chinese labs that led 2026's [text-to-video](https://metavert.io/text-to-video) rankings.

### Hailuo and the H3 Model

Hailuo AI launched in March 2024 and added video generation that September. MiniMax's current video model, MiniMax H3, which it also calls Hailuo 3.0, launched on July 31, 2026. It's a single "omni-modal" model that takes text, images, video and audio as input and generates clips of 4 to 15 seconds with 32 kHz stereo sound and speech in 11 languages. Output defaults to 768p, and an API-only module regenerates it at 2K. On Artificial Analysis's blind-vote arenas as fetched on October 20, 2026, H3 led the image-to-video board and placed fifth for text-to-video with audio, within a few points of Alibaba's [Wan](https://metavert.io/wan) 3.0 and ByteDance's [Seedance](https://metavert.io/seedance) 2.5. Other platforms also resell it: [Runway](https://metavert.io/runway)'s API lists MiniMax H3 among its third-party models.

### Open Weights, With Limits

MiniMax released part of H3 as [open weights](https://metavert.io/open-weight-models): the 33-billion-parameter H3-Base model, in two variants, at 768p. The context module and the 2K regeneration module stay behind the API. The MiniMax H3 Community License, dated August 2, 2026, defines its territory as worldwide "excluding the Excluded Territories", which it lists as "the European Union, the United Kingdom, the Republic of Korea and the United States of America". Users there must apply to MiniMax separately. Companies whose commercial products earn more than $20 million a year also need separate authorization. That's broader than Tencent's [HunyuanVideo](https://metavert.io/hunyuanvideo) license, which excludes the EU, UK and South Korea but not the US.

### Other Models and Disputes

Beyond video, MiniMax ships the M-series [language models](https://metavert.io/large-language-models) (MiniMax-M1 in June 2025, M2.5 in February 2026, M2.7 in March and M3 in June 2026) and the Speech-02 text-to-speech model (April 2025). Its video product has drawn legal challenges. In September 2025 Disney, Universal and Warner Bros. Discovery sued MiniMax, alleging that Hailuo infringed the rights to their characters. Those suits are part of a wider fight over training data and outputs in [generative video](https://metavert.io/generative-video).

### How It Compares

MiniMax competes most directly with the other Chinese video labs, Alibaba (Wan), ByteDance (Seedance) and Kuaishou ([Kling](https://metavert.io/kling)), and with [Google](https://metavert.io/google)'s [Gemini Omni](https://metavert.io/gemini-omni). H3's strengths are image-to-video quality, native stereo audio and partly open weights. Its weaknesses are short clips (15 seconds against Wan 3.0's and Seedance 2.5's 30) and a license that excludes the largest Western markets. For the choice between image-led and prompt-led generation, where H3 is strongest, see [Text-to-Video vs Image-to-Video](https://metavert.io/compare/text-to-video-vs-image-to-video).

## Related Topics

- [Text-to-Video](https://metavert.io/text-to-video) — How the models work and who leads in 2026
- [Text-to-Video vs Image-to-Video](https://metavert.io/compare/text-to-video-vs-image-to-video) — Where MiniMax H3 leads
- [Wan vs Seedance](https://metavert.io/compare/wan-vs-seedance) — The other Chinese leaders compared
- [Wan vs HunyuanVideo](https://metavert.io/compare/wan-vs-hunyuanvideo) — How license geography shapes open video models
- [Open-Weight Models](https://metavert.io/open-weight-models) — What open weights do and don't allow
- [Generative Video](https://metavert.io/generative-video) — The broader field
- [Kling (Kuaishou)](https://metavert.io/kling) — Another Chinese video model family
- [Alibaba Qwen](https://metavert.io/alibaba-qwen) — MiniMax's lead investor's own models

## Further Reading

- [MiniMax H3 on Hugging Face](https://huggingface.co/MiniMaxAI/MiniMax-H3) — model card, open and API-only modules
- [MiniMax H3 Community License](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE) — excluded territories and revenue threshold
- [MiniMax Group](https://en.wikipedia.org/wiki/MiniMax_Group) — Wikipedia; founding, products, listing and lawsuits
- [Artificial Analysis image-to-video leaderboard](https://artificialanalysis.ai/video/leaderboard/image-to-video) — where H3 ranked first
- [MiniMax's Hong Kong IPO](https://www.thestandard.com.hk/tech-and-startup/article/320794/Chinese-AI-firm-MiniMax-set-to-price-Hong-Kong-IPO-at-top-of-range-sources-say/) — The Standard; pricing at the top of the range