Zing-0.5: Toward Playable Worlds with Real-Time Joint Action and Text Control

Zing-0.5. Toward Playable Worlds with Real-Time Joint Action and Text Control. By Mingyang Chen and fifteen others, from SeedLeap dot ai. Published on the fifteenth of September. Its code repository is public and real: sixteen source files.

A generated world becomes a game only when a person can act in it and see the world respond. Zing-0.5 is built for that quality, which the authors call playability. Users explore generated worlds, influence the events unfolding in them, and respond to what happens, using the keyboard and online text together.

Zing-0.5 is an autoregressive world model with five billion parameters. It rests on three technical contributions: conditioning on action and text in one model, supervision at the scale of events for generation that proceeds incrementally, and interaction that is cheap enough to run in real time. Each contribution answers one requirement of play: being controllable in two ways at once, generating events as the world unfolds rather than all at the start, and responding fast and cheaply enough to feel live.

For the first, keyboard inputs that are aware of magnitude are combined with text instructions aligned in time, and with jointly annotated videos, so that navigation and event control are learned in the same sequence. For the second, a teacher that works on whole segments, trained on connected videos with multiple prompts, supervises a causal student that works block by block, through distribution-matching distillation.

For the third, generation in four steps is combined with streaming that preserves context. That supports inference at eight hundred and thirty-two by four hundred and eighty pixels, at twenty-four frames per second, at an estimated server rental cost of about zero point zero zero nine US dollars per stream-minute.

On one hundred and fifty-eight WBench Navigation cases, Zing-0.5 scores eighty-one overall and eighty-eight point five on consistency. A demonstration of joint control shows a text instruction changing an event while navigation carries on, without restarting generation. Changing what happens in the world is done in the middle of the stream, the way a player would expect, and exploration continues through the change.

The release includes the model weights, the inference code, and the Zing-SGLang serving implementation. The demonstration captures the point of the design: steering with the keyboard and directing events with text are not separate modes, and changing the story does not mean starting the world over.