Inside NVIDIA Cosmos | SIGGRAPH 2026 Keynote

Inside NVIDIA Cosmos | SIGGRAPH 2026 Keynote

🎙 NVIDIA Developer 👥 222K 📅 July 29, 2026 ⏱ 24 min 👁 6K 📄 keynote presentation 🧭 2026-08-13
Available in: English (current) Français

Keywords

world modelphysical AImixture of transformersaction modelsimulation

Summary

The keynote at SIGGRAPH 2026, presented by Ming-Yu Liu, VP of Cosmos Lab at NVIDIA, introduces Cosmos 3, a unified world foundation model for physical AI. The talk begins with a historical overview of NVIDIA’s contributions to graphics and AI, then addresses the data problem in physical AI, proposing that compute can generate data via world models. Cosmos 3 is an open omni-modal model based on a mixture of transformers architecture, combining an autoregressive tower for reasoning and a diffusion tower for generation, with shared multimodal attention. It supports multiple modalities including vision, text, audio, and action. Action is treated as a first-class modality, using geometric primitives to create a common action vocabulary across embodiments. The model supports forward dynamics, inverse dynamics, and policy modeling. NVIDIA releases three model sizes: Cosmos 3 Nano (16B), Super (frontier), and H (4B) for real-time on-device inference. The keynote also announces Cosmos Dreams, a neural closed-loop simulator for autonomous vehicles, demonstrated live on a single RTX Pro 6000 GPU. The presentation highlights partnerships and adoption, emphasizing the open platform with open models, datasets, and scripts.

181 words

Critical Evaluation

Value of the Information & Strength of the Argument

The value of the information is high for those interested in the latest developments in world models and physical AI. The keynote provides a clear explanation of the architecture and capabilities of Cosmos 3, including its unified approach to understanding, generation, and action. The argumentation is solid, supported by demonstrations and benchmark claims, though it is promotional in nature. The live demo of Cosmos Dreams on a single workstation GPU is compelling evidence of real-time performance. However, the presentation lacks detailed technical specifications and independent validation, relying on NVIDIA’s own benchmarks.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is moderate. The keynote is a product announcement, so it presents NVIDIA’s perspective without external peer review. The sources cited are NVIDIA’s blog and Hugging Face pages, which are primary but not independent. The title accurately reflects the content, and the presentation is well-structured. The claims about state-of-the-art results are based on NVIDIA’s own evaluations, which may be biased. No comments were provided for analysis.

174 words

Title / Content Match

The title accurately reflects the content: a keynote presentation about NVIDIA Cosmos at SIGGRAPH 2026.

Quality & Reliability

8/10

The keynote presents technical details of NVIDIA's Cosmos platform, including architecture and benchmarks, but is promotional in nature. Claims are not independently verified, and sources are limited to NVIDIA's own blog and Hugging Face pages.

Key Moments

Cited Sources

Concurring Sources

Contribution & Novelties

The keynote presents Cosmos 3 as a unified world foundation model that integrates understanding, generation, and action in a single model, which is a significant advancement over previous separate models. The introduction of geometric primitives for action representation enables cross-embodiment learning, and the release of Cosmos 3 H for on-device real-time inference is notable. The announcement of Cosmos Dreams for closed-loop simulation is also a novel contribution.

Pour aller plus loin :

102 words

Radar Profile

The radar profile shows high scores in quantity of information and technical level, reflecting the detailed technical content. Quality of information and global reliability are slightly lower due to the promotional nature and lack of independent verification.

Reliability 6/10