Odyssey released Odyssey-2 Max on Tuesday, the third and largest model in its Odyssey-2 world model series. The San Francisco startup, founded by former self-driving engineers from Cruise and Wayve, says the new model hits 58.52 on VBench 2's physics benchmark. NVIDIA's Cosmos-Predict2.5-14B scored 44.92 on the same test.
The numbers come from Odyssey's own blog post, so treat them with the usual caution. Odyssey-2 Max runs roughly 3x the parameters of Odyssey-2 Pro and used about 10x the training compute, trained across several hundred NVIDIA B200 GPUs in three stages.
Architecture-wise, it's an autoregressive diffusion transformer that predicts one frame at a time from prior states and user actions. CEO Oliver Cameron calls it "pretrained physical intelligence," likening the moment to GPT-2 before ChatGPT. Whether that analogy holds depends on what partners actually build.
Odyssey-2 Max is available in private beta for partners working in robotics, gaming, simulation, defense, and interactive systems. No pricing disclosed. The earlier Odyssey-2 Pro has been live via API since earlier this year.
Bottom Line
Odyssey-2 Max scored 58.52 on VBench 2's physics benchmark, compared to 44.92 for NVIDIA's Cosmos-Predict2.5-14B, though the numbers are self-reported.
Quick Facts
- VBench 2 physics score: 58.52 (company-reported)
- NVIDIA Cosmos-Predict2.5-14B score on same benchmark: 44.92
- 3x parameter count vs Odyssey-2 Pro
- 10x training compute vs Odyssey-2 Pro
- Trained on several hundred NVIDIA B200 GPUs
