AI startup Odyssey just dropped two models in two days: Starchild-1 and Agora-1.
Starchild-1 tackles real-time audiovisual interaction. Most video generation models — take DeepMind’s Veo — work offline. You hit generate, and the video’s length and trajectory are locked in. No mid-stream edits allowed.
Starchild-1 uses causal distillation and an asynchronous KV cache to separate how audio and video context are stored. That means you can shout out a command, type something, or make a gesture at any point, and the model instantly adjusts what comes next — both the visuals and the environmental sound.
Then came Agora-1, which extends that real-time interaction to multiplayer. It supports up to four human or AI players battling it out in a single round of GoldenEye 007.
Earlier attempts at multi-agent interaction were shaky. Multiverse used a split-screen state representation; Solaris crammed every participant into one Transformer sequence. As the player count grows, the context window balloons. If players split up, the shared world state falls out of sync.
Agora-1 separates the underlying game logic from rendering. It first calculates each player’s exact position and health, maintaining a perfectly consistent shared state. Then a DiT model reads that state and renders a first-person view for each player individually.
Together, these two releases make Odyssey’s direction clear: transform world models from one-shot video generators into interactive engines that can hear, see, and run multiplayer sessions.
TechCrunch previously reported that Odyssey was founded by two entrepreneurs with backgrounds in autonomous driving. The startup has raised $27 million to date.