Singapore Video Startup Founded By Tencent’s Former AI Head Bets Big On World Models

Rarely does a startup emerge not just with ambition, but with a blueprint that challenges the hierarchy of the AI arms race—especially when going headfirst into one of the most resource-heavy domains: AI-generated video and world modeling. Video Rebirth isn’t backed by a tech titan, nor does it wield infinite compute, yet in May, its Bach model landed at No. 6 on the Artificial Analysis text-to-video leaderboard—highest among independent startups and, crucially, the cheapest per minute in the top 10. That’s not luck. That’s architecture.

What’s striking isn’t just their efficiency, but their focus. Liu Wei, former Tencent scientist and now CEO of Video Rebirth, didn’t jump into video generation to churn out flashy short clips and call it a day. From the start, video was just the scaffold. The real goal? Building a world model—one that simulates physics, understands cause and effect, and anticipates outcomes in real time, not just for entertainment, but for robotics, mobility, and simulation at scale.

Their approach is ruthlessly pragmatic. By training on fewer, higher-quality 720p clips—licensed films, music videos, and in-house footage—they cut training costs dramatically. More clever still, they split prompt understanding from visual generation, reducing computational overhead. Their proprietary multi-step sampling loss technique reduces generation steps by predicting and correcting errors early, accelerating inference up to 10 times. This isn’t incremental improvement—it’s a recalibration of the entire engine.

And while OpenAI quietly shelved Sora despite Disney’s involvement, citing unsustainable inference costs, Video Rebirth claims significantly lower per-clip generation costs. That kind of efficiency, combined with physical realism—gravity, object interactions, consistent lighting—positions them uniquely for enterprise. Advertisers get consistent product rendering. Filmmakers get believable motion. Game studios get stepping stones toward interactive environments.

Partners like Hyundai’s ZER01NE and CJ Group’s Hiven aren’t just investors—they’re strategic visionaries betting on hyper-realistic digital worlds for mobility training and content. When Hyundai touts Video Rebirth as a ‘key partner for the future of mobility,’ it’s not PR. Boston Dynamics and autonomous vehicles need simulated worlds that behave like our own. That’s the prize.

Liu’s foresight in leaving Tencent to build on 'physical AI'—a space he saw as wide open while LLMs consolidated—proves prescient. By 2026, they aim to launch Olympus, a world model with real-time 3D simulation plus environmental sound, moving beyond Genie 3’s early capabilities. If they deliver, this tiny Singapore-based team of 30 won’t just be punching above their weight—they’ll be defining the next weight class.

This is one of those rare moments where discipline, architecture, and vision converge. The world model race is still in its infancy, but Video Rebirth isn’t waiting for permission.

Curious how a lean team outmaneuvers giants in AI video? The answer’s not in scale—it’s in smarts. Read the full piece to see how they’re building the future, one simulated second at a time.