ByteDance is reportedly building an AI system that generates navigable video worlds in real time, with founder Zhang Yiming said to be personally driving the project and an unconfirmed release window of next month. Bloomberg broke the account, citing unnamed people close to the plans, and one of those people told Bloomberg the schedule is not settled and could move. The Next Web, which relayed the story, said it could not independently verify the details and that a ByteDance spokesperson did not answer a request for comment.
The reported system would sit on top of Seedance, the company’s existing video generation engine, and would aim to give users interactive three-dimensional scenes for livestreams, short dramas and games rather than static clips. Pico, the extended reality hardware arm ByteDance owns, is reportedly the target: environments that react as a Pico wearer speaks and moves.
Bloomberg’s sourcing puts the target performance at roughly 20 frames per second with about 50 milliseconds of lag, computed remotely rather than on the headset itself. That distinction is the entire bet: keeping the heavy computation in the cloud is what would let Pico ship cheaper hardware instead of chasing Vision Pro grade optics. Real time is the genuinely hard part of this category, and every system announced so far, from World Labs’ Atlas to Runway’s GWM Worlds 2, has had to compromise on frame rate, resolution or how long a session holds together before it breaks. The number worth tracking once ByteDance actually ships anything is not that it built a world model at all, but what frame rate and session length it can sustain once real users are pushing on it live.
36Kr reported separately earlier in the year that world models sit at the top of ByteDance’s internal AI priorities for 2026, above holding Seedance’s video lead, above better coding tools, and above making Doubao earn revenue. Those sources said the data budget for world model training ran several times higher than what rivals were said to be spending, with an internal goal of matching or beating Google’s Genie, the DeepMind system that renders a walkable version of Street View, by year end.
Bloomberg presents Zhang as personally involved rather than merely signing off, and situates him with researchers including Fei-Fei Li and Yann LeCun, whose case is that grounding a model in the physical and visual world beats relying on language by itself. That is Bloomberg’s characterization of his involvement, sourced to people it did not name, not a confirmed title or mandate, and no outlet in this chain has independent confirmation beyond that sourcing.
That same 36Kr account put ByteDance’s internal results roughly ten percentage points short of the best systems anywhere at that point, so shipping next month would represent a jump rather than a scheduled step, which is one more reason to hold the date loosely.
None of this is happening on a shoestring. ByteDance reportedly secured a $30 billion loan last week and has discussed capital spending of up to $70 billion on its AI build-out, the kind of budget that can absorb a slipped deadline without much notice.
AI Insiders has tracked this category taking shape all year: World Labs shipped Atlas, Runway shipped GWM Worlds 2, and the World Labs founders have argued in public conversation, including with a16z, that persistent, interactive worlds are the step past chatbots. A ByteDance entry, if it arrives anywhere near the reported spec, would be the first to pair that ambition with an existing headset line and a user base already trained to expect instant video playback.
Anyone evaluating XR or generative-world tools this quarter should treat next month as a possible date, not a launch window, and instead watch for ByteDance to publish an actual frame rate and session-length figure that a Pico headset can hold during a live demo. That number, not the announcement, is what will separate this attempt from the rest of the field’s compromises.
The Next Web reported this on September 7, 2026, citing Bloomberg’s sourcing on ByteDance’s plans.
