Alibaba launches HappyOyster, an interactive playable world model

For the past two years, generating AI video has followed a familiar routine: you type a prompt, wait for the render queue, and pray that the resulting five-second clip looks somewhat close to what you wanted. If a character walks the wrong way, your only option is to rewrite the prompt and roll the dice again.

Alibaba Cloud introduced HappyOyster 1.0, an open world model designed to flip that passive dynamic on its head. Instead of spitting out static video files, it builds interactive environments in real time.

The project originally entered closed beta testing back on April 16. Two months later, Alibaba has officially polished the system into a 1.0 release that anyone can try online.

From watching video to playing it

HappyOyster 1.0 relies on a native multimodal architecture capable of joint audio and video generation. According to Alibaba, the model has been trained on state transitions in the physical world, allowing it to deduce cause-and-effect chains and keep characters and environments consistent over longer stretches.

Rather than forcing you into a rigid "prompt, wait, export" loop, the model accepts user input continuously while the scene is playing out.

The web interface offers two primary interaction modes. The first is Directing mode, where you can interrupt a narrative and rewrite its direction on the fly. You can chat with virtual companions or pause a short drama mid-scene to rewind and branch the storyline elsewhere.

The second mode, Adventure, basically turns generated video into a playable sandbox. Users can hit the WASD keys on their keyboard to walk around, jump, hide, and engage in simple combat. It lets you explore environments ranging from polar ice sheets to deep oceans, with movement mechanics that include skateboarding, wingsuit gliding, and horseback riding.

Alibaba Cloud noted that daily login credits will be dropped on its official website through July 17 to encourage experimentation.

Why it matters

In our view, HappyOyster represents an important shift in generative media: moving from passive video viewing to real-time interaction.

While most frontier models are still racing over higher resolutions and longer video clips, the real holy grail has always been responsiveness. By integrating keyboard controls and causal physics into the generation loop, Alibaba is testing whether neural networks can effectively bypass traditional graphics rendering pipelines for simple interactive worlds.

For creators and game designers, this dramatically cuts down the time required to prototype branching stories or mini-games. You no longer need to script every physics interaction by hand when the model can intuit physical feedback on the fly.

The big question mark remains stability. Handling real-time inputs without breaking visual coherence or hallucinating the scenery away after several minutes of open exploration is notoriously difficult. Still, making an AI-generated world truly playable with standard keyboard controls is a massive step in the right direction.