Odyssey Opens Free Preview of Its 14-Billion-Parameter World Model

Odyssey opened a free browser preview of Odyssey-3, a 14-billion-parameter autoregressive diffusion transformer world model that generates interactive first-person and third-person environments from text prompts at 832×480 resolution, with Odyssey-3 Pro producing 1280×720 video. Odyssey reported a 66.1 Physics-IQ Verified score for Odyssey-3 Pro using best-of-eight sampling, while independent coverage reported 63.37 across four runs without that selection method. The preview does not expose Odyssey-3's robot, vehicle or drone controllers, and no downloadable weights, local hardware requirements or public Odyssey-3 API pricing were published.

Odyssey Opens Free Preview of Its 14-Billion-Parameter World Model - Odyssey-3 is available through a free browser preview that generates interactive first-person and third-person environments from text prompts. <sup \ 1\ https://odyssey.systems/meet-odyssey-3 </sup - The base model is described as having 14 billion parameters and producing 832×480 video; Odyssey-3 Pro supports 1280×720 output. <sup \ 2\ https://the-decoder.com/odyssey-3-is-a-new-generative-world-model-that-you-can-try-for-free/ </sup - Odyssey-3 Pro posted a 66.1 Physics-IQ Verified score using best-of-eight sampling. A four-run, single-generation figure reported by independent coverage was 63.37. <sup \ 3\ https://the-decoder.com/odyssey-3-is-a-new-generative-world-model-that-you-can-try-for-free/ </sup - Odyssey’s robot, driving and drone results are company demonstrations built around specialized controllers; no independently reproducible robotics evaluation package has been published. <sup \ 4\ https://odyssey.systems/introducing-odyssey-3 </sup Odyssey has opened a free public preview of Odyssey-3, a 14-billion-parameter world model that generates interactive environments from text prompts and updates them as users move, change viewpoints or introduce events. The California startup is presenting the system as both a browser-based research experiment and a foundation for physical AI. \ 1\ https://odyssey.systems/meet-odyssey-3 \ 2\ https://the-decoder.com/odyssey-3-is-a-new-generative-world-model-that-you-can-try-for-free/ The preview demonstrates interactive world generation. It does not provide public access to Odyssey-3’s robot, vehicle or drone controllers. Odyssey’s launch post invites physical-AI developers to request access, but the developer portal currently describes access to Odyssey-2 Max rather than publishing Odyssey-3 API documentation or pricing. \ 1\ https://odyssey.systems/meet-odyssey-3 \ 5\ https://developer.odyssey.ml/ A playable model with undisclosed local requirements Odyssey-3 is an autoregressive diffusion transformer. It predicts how a scene changes from previous video frames, user actions and other controls, then generates additional frames as the interaction continues. The preview supports first-person and third-person navigation, along with independent camera movement. \ 1\ https://odyssey.systems/meet-odyssey-3 The base model is listed at 14 billion parameters and produces 832×480 video, according to reporting that reviewed Odyssey’s benchmark submission. Odyssey-3 Pro generates 1280×720 video. Odyssey says its real-time variant uses distillation and fewer diffusion steps to reduce the compute required for interaction. \ 2\ https://the-decoder.com/odyssey-3-is-a-new-generative-world-model-that-you-can-try-for-free/ \ 6\ https://odyssey.systems/meet-odyssey-3 The reviewed public materials do not state the hardware needed to run Odyssey-3 locally, and no downloadable weights or detailed model card were located. The hosted preview is available through Odyssey’s website; the developer portal does not currently expose public Odyssey-3 pricing, rate limits or technical API documentation. \ 1\ https://odyssey.systems/meet-odyssey-3 \ 5\ https://developer.odyssey.ml/ The benchmark result is narrower than the headline Odyssey says Odyssey-3 Pro achieved a 66.1 score on Physics-IQ Verified’s video-to-video benchmark, which asks models to predict the continuation of videos showing experiments involving fluid dynamics, optics, solid mechanics, magnetism and thermodynamics. \ 1\ https://odyssey.systems/meet-odyssey-3 \ 7\ https://arxiv.org/abs/2606.18943 That number needs qualification. It uses a best-of-eight selection procedure: eight candidate continuations are generated and a preferred result is selected. Independent coverage reports a 63.37 score for Odyssey-3 Pro across four runs without that selection method. The benchmark documentation distinguishes one-run leaderboard entries from the multi-run evidence expected for a state-of-the-art claim. \ 3\ https://the-decoder.com/odyssey-3-is-a-new-generative-world-model-that-you-can-try-for-free/ \ 7\ https://arxiv.org/abs/2606.18943 Odyssey also reports leading results on WorldMark, a benchmark that measures control-following, visual quality and consistency over time. Its evaluation places Odyssey-3 first in three of four categories, but third in the first-person real category, behind Lyra 2.0 and AlayaWorld. Odyssey produced the evaluation using WorldMark’s captions and scoring approach; it is not an independent audit of the company’s model. \ 1\ https://odyssey.systems/meet-odyssey-3 \ 3\ https://the-decoder.com/odyssey-3-is-a-new-generative-world-model-that-you-can-try-for-free/ Robotics demonstrations still need outside testing Odyssey says the same pretrained model can support robot arms, humanoids, vehicles, drones and game-playing agents. Each application uses an additional action decoder or policy that translates the model’s learned visual representations into machine-specific commands. \ 4\ https://odyssey.systems/introducing-odyssey-3 The company reports robot-arm recovery behaviors after missed grasps, humanoid control developed with Swiss robotics company Flexion, autonomous driving on Indian roads after training with 20 hours of simulated driving data, and drone navigation in a simulated indoor environment. Odyssey also demonstrated an AI playing GTA V after roughly two hours of game footage. \ 4\ https://odyssey.systems/introducing-odyssey-3 These examples test more than the public preview, but they do not establish that one general-purpose model can reliably control different machines outside controlled demonstrations. The disclosed drone work is simulated, and the driving comparison comes from Odyssey’s own evaluation of simulation-trained and real-footage-trained policies. \ 4\ https://odyssey.systems/introducing-odyssey-3 Independent discussion of the launch has identified the absence of a complete, independently reproducible robotics evaluation package. More broadly, experts quoted by Axios have warned that learned simulators cannot be assumed to capture every rare hazard and require substantial real-world testing before they are used in safety-critical systems. \ 8\ https://www.axios.com/2026/02/25/ai-waymo-robotaxis-av What is available now For the public, Odyssey-3 is currently a hosted research preview for generating and exploring virtual environments. The company is inviting physical-AI developers to request access, but the reviewed launch materials do not include public Odyssey-3 weights, API pricing or a complete model card. \ 1\ https://odyssey.systems/meet-odyssey-3 \ 5\ https://developer.odyssey.ml/ That leaves Odyssey-3 in an early research phase. The preview makes the model’s interactive generation easy to inspect, while the more consequential claims about sim-to-real transfer, recovery behavior and safety remain dependent on company demonstrations and benchmark submissions that outside teams have not yet reproduced. Further sources The stories that matter, in one email. Free — unsubscribe anytime.