Chinese AI developer MiniMax released a video generation model called H3 on Friday, and its Hong Kong-listed shares rose 17.4% to HK$230.0 in early trading. The model makes clips of up to 15 seconds at 2K resolution with sound produced alongside the picture, and it can also edit existing footage and copy motion between clips when given reference material in other formats.
The size of the share move says as much about the stock as about the model. MiniMax listed in Hong Kong in January at HK$165, doubled on day one, and touched HK$1,330 in March. Then it fell. A July 9 lockup expiry released about 49% of its shares to trade at once, JPMorgan cut its price target twice in one week, and the shares closed at HK$193.1 on July 20. Even after Friday's jump, they sit more than 80% below the March high.
So H3 is the first evidence in months that MiniMax can move its own stock with a product rather than with share supply. It enters a market others lead. On Artificial Analysis, which ranks video models by blind side-by-side viewer votes, 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.
One model instead of a dozen specialists
Until now, generating video with AI has meant picking a specialist. One model turns text into video. A different one animates a still image. Another swaps a character out of a finished shot. Voice, sound effects and music each come from their own systems. Building a 15-second ad means moving files between four or five tools and hoping the character's face survives the trip.
MiniMax says H3 collapses that. The company writes that "the first principle guiding H3's development was unifying and generalizing across tasks," meaning one model that takes any mix of pictures, clips and audio as reference and works out from a plain-language sentence what to do with them. Its own example asks the model to borrow a camera move from one video, put it on a face from an image, and sing in a voice from an audio file.
For readers outside the field, the claim to judge is not picture quality. It is whether the tool count in a production pipeline drops from five to one, which is what MiniMax is selling to advertisers and online retailers rather than to filmmakers.
The price claim and the open-weights promise
MiniMax says 2K video from H3 costs less than one-third of what mainstream rivals charge per second, and that its 768p output costs under half what competitors charge for 720p. Those are company figures, not independent tests.
The bigger commitment is distribution. MiniMax says it will publish H3's weights within days, subject to law and regulation. Weights are the trained numbers that make a model work, so publishing them lets anyone download H3 and run it on their own hardware without paying for access. Video has stayed mostly closed while Chinese labs opened up their text models. MiniMax also says H3 was built to run on several Chinese-made chips.
That fits a pattern that has been reshaping the wider market. DeepSeek anchors the low end on price, Alibaba's Qwen family has the widest developer following, Moonshot released the 2.8-trillion-parameter Kimi K3 on July 17 with weights to follow, and MiniMax already publishes its M3 language model. The effect on American labs shows up in routing data rather than benchmarks. Chinese models' share of tokens sent by U.S. firms through the OpenRouter marketplace peaked at 46% by mid-2026, against an 11% average over the prior year, and open Chinese models run 60% to 90% cheaper than leading OpenAI and Anthropic systems. Volume is not revenue, and closed U.S. models still earn far more per token on premium work. But the price floor is being set in China, and H3 extends that from text into video.
What has not been published
H3's size is not public. Without a parameter count, nobody outside the company can say whether a downloaded H3 runs on a workstation or needs a rack of accelerators, which decides whether the open-weights promise means much in practice. Pricing, rate limits and licence terms for commercial output are also unpublished.
Nor is there an independent quality number. H3 had no entry on the Artificial Analysis video leaderboards on Friday, so the only side-by-side comparisons available are MiniMax's own demonstration clips, chosen by MiniMax. Early testers suggest its edge is cost and audio rather than high-motion camera work, where rivals still lead.
MiniMax has promised a full technical report and says it plans to fold in its language models next. Investors who bought Friday's move are betting the weights arrive on schedule and that a cheaper, more general model wins volume. The company has a few days to show it.