MINIMAX’S H3 LICENSING CURBS REVEAL THE GROWING COST OF TAKING CHINESE AI VIDEO MODELS GLOBAL
MiniMax introduced H3 as a general-purpose, multimodal system capable of processing combinations of text, images, video and audio. The model can produce clips lasting between four and 15 seconds at up to 2K resolution, with 24-frame-per-second video and native stereo audio. It supports text-to-video, first-and-last-frame generation and reference-based creation using multiple images, videos and audio clips. H3 can also edit existing content, transfer movement between videos and generate dialogue in several major languages. The publicly released model has approximately 33 billion parameters, enabling developers to download, customise and deploy its underlying weights instead of depending entirely on MiniMax’s proprietary platform. This approach gives the company a way to attract researchers and application developers in a video-generation market that has traditionally been dominated by closed systems.
The release is not geographically unrestricted, however. The standard H3 Community License excludes users in the US, EU, UK and South Korea, requiring companies in those regions to apply for formal permission. MiniMax has said that it will review each applicant’s deployment scenario and determine whether appropriate content controls and legal safeguards are in place. At the same time, customers in these markets can continue accessing H3 through MiniMax’s paid API or authorised third-party services. This creates an important commercial distinction. When developers possess model weights, they can modify the system, remove safeguards or deploy it without continuous supervision. When they use an API, MiniMax retains control over the infrastructure and can enforce restrictions involving copyrighted characters, deceptive likenesses, minors and other unsafe content. The licensing policy therefore channels some of the world’s most valuable AI markets toward a more controllable and potentially recurring-revenue service.
Copyright litigation is the most immediate reason for this caution. Disney, Universal and Warner Bros. Discovery sued MiniMax in a California federal court in September 2025, alleging that the company used protected material to develop Hailuo AI and enabled users to generate unauthorised images and videos featuring characters from franchises including Marvel, Star Wars and Despicable Me. The studios also accused MiniMax of using recognisable characters to market Hailuo as a “Hollywood studio in your pocket.” MiniMax attempted to have the case dismissed, arguing partly that a US court lacked jurisdiction because the company conducted its business from China. In May 2026, the judge rejected that request after finding evidence that Hailuo was offered to US users and concluding that the studios had presented legally plausible claims. The decision was procedural and did not establish that MiniMax infringed copyright, but it allowed the case to proceed and increased the financial and operational uncertainty surrounding unrestricted US deployment.
The restrictions arrive as MiniMax is trying to convert technological progress into sustainable commercial growth. H3 is aimed at advertising, e-commerce, film production, video games and product design, where faster and cheaper content generation could create significant enterprise demand. MiniMax has claimed that producing 2K video with H3 costs less than one-third as much as mainstream competing products. Independent benchmark results have been more nuanced: H3 has performed strongly in video-editing evaluations but has trailed selected rival models in some text-to-video and image-to-video tests. Competition is also intensifying as ByteDance advances its Seedance series and Kuaishou develops Kling. Open weights and lower prices can accelerate adoption, but they may also compress margins and make it harder for MiniMax to recover the substantial computing and research expenditure required to train increasingly capable models.
The financial stakes are considerable because international business is already central to MiniMax’s growth story. The company’s 2025 revenue rose 159% to US$79 million, with more than 70% generated outside China. Consumer subscriptions and enterprise-platform services both expanded rapidly, but MiniMax remained loss-making. It reported a US$1.87 billion net loss, although much of that figure resulted from changes in the value of financial instruments rather than ordinary operating expenses. MiniMax raised approximately HK$4.8 billion through its January 2026 Hong Kong initial public offering, and its shares more than doubled on their first trading day as investors focused on the potential of Hailuo AI, Talkie and its multimodal foundation models. H3’s global adoption could support the valuation by expanding usage and enterprise revenue, while an adverse legal judgment, licensing costs or prolonged restrictions in major overseas markets could weaken the international growth assumptions embedded in the company’s investment case.
MiniMax’s strategy ultimately reflects a broader transition in China’s AI sector. Chinese developers initially gained international attention by releasing capable models at lower prices and with more accessible weights than many US competitors. Generative video now demonstrates the limits of that playbook because it can reproduce characters, faces, voices and cinematic styles in ways that create immediate legal exposure. MiniMax is attempting to preserve openness where regulatory risks are manageable while retaining central control in the jurisdictions most likely to impose liability. The result is a hybrid model rather than unrestricted open source: downloadable technology for much of the world, case-by-case licensing in sensitive markets and globally available paid access through company-controlled infrastructure. Its success will depend not only on H3’s technical quality, but also on whether MiniMax can turn compliance into a competitive advantage without undermining the international developer adoption that its open-weight strategy was designed to achieve.











