Real-time AI video generation advances with MiniMax H3 Max and Visko Orbis

According to a report by QbitAI, MiniMax and fal jointly developed H3 Max by optimizing the open-source H3 video model for real-time inference. The model generates a five-second 768p video in under three seconds, achieving roughly 35 times the throughput of the original H3. H3 Max also ranks first for image-to-video on both the Artificial Analysis and Design Arena leaderboards.

H3 Max's speed has spawned applications that were previously impossible, the report said. Rehan Sheikh, an engineer at fal, integrated H3 Max into a live streaming platform to create a "cross-dimensional television" channel, where all visuals and audio are generated in real time by the model. A clip of the stream he posted to X has been viewed 5.4 million times. Independent developer Pieter Levels built a live AI streaming site where viewers can request content on demand, and fal engineer Alex Koumpas launched a stream that presents a continuous story, with viewers able to intervene via chat commands. Within a week, the three created AI live streams independently, without coordinating.

The report also noted that H3 Max is part of MiniMax's broader open-source push. The original H3 model, open-sourced in late July, had been downloaded 24 million times in three weeks and spawned over 300 derivative models. On Tuesday, MiniMax shares rose 16.18 percent in Hong Kong to HK$349, after Morgan Stanley maintained an overweight rating on the company with a target price of HK$900.

In a separate development, Visko Platform Inc. said it has raised $10 million in a pre-seed round from Llama Ventures and opened public access to its first foundation model, Orbis, according to SiliconANGLE. Visko is an AI research lab focused on persistent world models. Orbis is built on a new "live model" paradigm that runs continuously, generating and streaming interactive visual worlds in real time. Unlike conventional query models such as Sora, which generate short static clips, Orbis uses a learned dynamic system governed by neural stochastic differential equations, maintaining a persistent internal state that anticipates motion and retains object detail. Users can intervene at any moment with new prompts, and the output updates immediately without stopping the stream.

Orbis streams interactive video at 4K resolution and 24 frames per second, and can continue generating worlds for hours without observable degradation in color or quality, the company said. Visko published a technical report evaluating Orbis on the DOVER and VideoAlign benchmarks, where it outperformed competitors including Odyssey-2, and ranked first in a human preference Arena study for long-form performance. Founder and CEO Qing Yin said long-term coherence has been the biggest challenge in AI video generation, and that Orbis was built to hold together physically, visually and narratively for as long as users want to stay inside it.

Dr. Michael I. Jordan, a professor at the University of California at Berkeley and an advisor to Visko, said the startup delivers a state-of-the-art architecture with a meaningful technological lead beyond conventional short-clip video generation, and the core capability could support applications ranging from robotics and physical simulation to gaming and creative media. Constellation Research analyst Holger Mueller called Visko's work encouraging, noting that generating world models dynamically rather than statically means they never have to stop, though cost may be an issue for continuous generation.