Shares in Chinese AI company MiniMax rose 17.4% on the Hong Kong stock exchange this Friday, peaking at HK$230. The catalyst was the release of a new video generation model named H3, which appears to have granted some relief after a turbulent few months for the firm.
MiniMax went public this January at a price of HK$165 and enjoyed a meteoric 700% rise to its March peak. Since then, however, it has dropped considerably. JPMorgan cut its price target for the stock twice in July, following a lockup expiry that flooded the market with roughly 49% of the company’s shares at once.
H3’s unified generation workflow
MiniMax’s H3 is capable of producing 2K clips of up to 15 seconds in length, with accompanying audio. It can also copy motion between clips and edit existing material. However, its USP is less about the model’s technical capabilities and more about its streamlined workflow. In a post accompanying the reveal, the company wrote that "the first principle guiding H3's development was unifying and generalizing across tasks."
AI video production has previously required a whole suite of specialized tools. A piece of raw footage is first generated using a text-to-video or image-to-video model, then passed down a pipeline of additional models for editing and refinement. These tweak character movements, replace faces, add sound effects, or perform any number of other specialized tasks. Maintaining consistency throughout this pipeline requires careful control by the creator.
MiniMax has deliberately sought to merge all of these capabilities into a single model, allowing users to combine multiple information sources into one prompt. To showcase what this looks like in practice, the company walked through an example wherein it asked the model to take the camerawork from one clip, insert a face from a still image, and synchronize the vocals from an audio file with the character’s lips.
Advanced video editors may prefer the greater level of control that comes with a suite of specialized tools. However, MiniMax is promoting this unified workflow to online advertisers and retailers rather than filmmakers, promising the ability to spin up marketing materials with limited friction.
The price claim and the open-weights promise
The AI video generation market is highly competitive. Artificial Analysis, the independent model ranking platform, performs blind tests for new models by giving a panel of viewers side-by-side comparisons and asking them to vote. At present, ByteDance's Seedance 2.0 tops text-to-video with audio at 1,214 Elo, ahead of Kuaishou's Kling 3.0 and Google's Veo 3.1.
Alongside ease of use, MiniMax also aims to compete against these models on price. The company claims that its model generates 2K video for under one-third of the cost of its competitors, while its 768p clips come in at less than half of 720p footage from mainstream models.
Users with the hardware to run the model locally will be able to do so for free, as MiniMax plans to release H3’s model weights in the coming days - unusual for a video-generation model. Many Chinese firms have already opened access to their text models, which has led to a sharp rise in the number of U.S. firms utilizing them.
Chinese models accounted for 46% of all tokens sent through the OpenRouter marketplace by U.S. firms during a mid-2026 peak, compared to an average of 11% in 2025. These models can reportedly run as much as 90% cheaper than leading closed models from U.S. firms.