Odyssey-2 Max

Odyssey-2 Max Introduction

Odyssey-2 Max is Odyssey's largest general-purpose world model, designed for real-time interactive simulation with advanced physical accuracy. Using autoregressive next-state prediction, it achieves long-horizon stability and open-ended interaction, outperforming competitors like Sora and Veo in physics benchmarks.

Key Features

Autoregressive next-state prediction for high-fidelity simulation

Real-time interactive simulation with over 120 seconds of stable generation

State-of-the-art physical accuracy (VBench 2.0: 58.52, PAI-Bench: 93.02)

Scalable architecture with 3x parameters and 10x training compute over predecessor

Open-ended interaction allowing user actions to evolve the simulated world

Use Cases

Robotics companies for training and testing control policies in simulated environments

Game developers for creating realistic, interactive virtual worlds

AI researchers studying physical dynamics and long-horizon predictions

Autonomous vehicle teams for simulating edge-case driving scenarios

Industrial designers for prototyping physical systems and interactions

Why Startups Use It

Startups leveraging AI for physical simulations can use Odyssey-2 Max to accelerate R&D without costly real-world data collection. It enables realistic testing of robots, autonomous systems, and interactive experiences at a fraction of the time and cost, giving startups a competitive edge in innovation.

Frequently Asked Questions

What makes Odyssey-2 Max different from traditional video generation models?

Odyssey-2 Max is a causal world model that predicts future states interactively, unlike bidirectional video models which cannot condition on user actions or simulate open-ended futures.

Is Odyssey-2 Max available for public use?

It is currently available in private beta for select partners in robotics, gaming, and other sectors. You can request access via the Odyssey website.

What architecture does Odyssey-2 Max use?

It is based on the Autoregressive Diffusion Transformer architecture, employing continuous flow matching and few-step denoising distillation for real-time inference.

How does Odyssey-2 Max compare to other world models on physics accuracy?

It achieves top scores on VBench 2.0 (58.52) and PAI-Bench (93.02) physics subtasks, outperforming models like Odyssey-2 Pro and NVIDIA's Cosmos-Predict2.5-14B.

Can Odyssey-2 Max be used for long-duration simulations?

Yes, it can generate simulated footage continuously for over 120 seconds without collapsing, demonstrating superior long-horizon stability.