MiniMax Surges Over 19% as H3 Max Video Generation Speed Exceeds Playback Speed
nashnova research
MiniMax jumped 19% to HK$358 on turnover of HK$5.8 billion after its H3 Max model generated video faster than real-time playback for the first time — turning AI video from a one-clip tool into infrastructure that can feed a continuous livestream.
What does "faster than playback" actually mean?
H3 Max, post-trained and inference-optimized by fal on MiniMax's H3 open-weight release, generates a full 5-second, 768p audio-video clip in under 3 seconds.
This means → while the previous clip is still playing, the next one is already queued — the model can now plug into a non-stop livestream.
In plain terms = AI video used to be "click, wait, get one clip." Now it can broadcast continuously, like a TV station.
How much faster, and how was it done?
fal reports H3 Max throughput at roughly 35× the MiniMax H3 official endpoint; a 15-second clip takes about 15 seconds to generate.
The speedup comes from co-optimizing model and inference stack: fal added new data on top of the open weights and ran verifiable reinforcement learning — a training method where the model checks its own output quality and improves accordingly.
fal also adapted the model to its own inference infrastructure. This reflects a viable path for AI-video deployment: open weights + third-party inference optimization.
Who is already using it, and for what?
fal engineer Rehan Sheikh connected H3 Max to Twitch and built a "cross-dimensional TV" show with no fixed schedule — the model generates visuals and audio continuously and streams them live.
The related videos have reached nearly 5.3 million views on X.
This means → the addressable user has shifted: from "someone who needs to produce a video clip" to "a content product that needs a continuous stream of visuals."
What do the benchmarks and specs tell us?
H3 Max ranks first on both the Design Arena image-to-video leaderboard and the Artificial Analysis image-to-video leaderboard.
It supports text-to-video and image-to-video, outputting 480p or 768p, 5–15 seconds, 24 fps full audio-video.
In plain terms = topping both leaderboards means quality and instruction-following held up — the speed gain did not come at the expense of output quality.
What do brokers say, and what is the next proof point?
Huatai Securities expects AI-model commercialization to accelerate over the next four to six months, noting an improving competitive landscape in MiniMax's segment.
Huatai forecasts faster product launches in H2 to drive revenue growth; M3 Pro could launch in September–October with a parameter scale potentially reaching 3 trillion.
Guosheng Securities and Guojin Securities both maintain a "Buy" rating. This means → the market's next checkpoint is clear: whether real-time livestream use cases can generate commercial revenue at scale — the core test of MiniMax's monetization delivery.
市场有风险，内容仅供研究参考，不构成投资建议。