English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

NVIDIA Cosmos 3: A Multimodal Foundation Model for Physical AI

Forum topic · 小凯 · 2026-08-08

Summary

NVIDIA unveiled Cosmos 3 at Computex 2026, repositioning its world model line from video generation toward a unified multimodal foundation for physical AI. The release ships in three sizes: Cosmos-3 Super (64B), Cosmos-3 Nano (16B), and Cosmos-3 Edge (4B). Its Mixture-of-Transformers dual-tower architecture processes video and image tokens in one branch and text, action, and sensor signals in another, fused via cross-attention. Training used about 1.3 billion data points spanning 85M hours of simulated video, 450M action sequences, 280M text annotations, and 120M sensor readings. The NVIDIA Open Model License v1.1 permits commercial derivative work but excludes military use and deepfake synthesis. Cosmos 3 underpins GR00T, Isaac Sim, Isaac Lab, and Omniverse workflows, and differentiates itself from Genie 3, Veo 3, and Sora 2 by treating action tokens as a first-class output alongside video.

Overview

NVIDIA formally released Cosmos 3 at Computex 2026 (May 31 Beijing time), with a follow-up NVIDIA Blog post on August 6 titled *Open World Models Power the Next Frontier of Physical AI*. The release reframes the Cosmos line: rather than a video generator, Cosmos 3 is positioned as a multimodal foundation for physical AI, unifying vision, action, text, and sensor reasoning within a single set of weights.

Three Model Sizes

Cosmos 3 ships as a family of three checkpoints:

  • Cosmos-3 Super (64B) — flagship for physics simulation, robot policy learning, and autonomous-driving scene generation.
  • Cosmos-3 Nano (16B) — tuned for edge and embedded deployment with latency-sensitive use cases.
  • Cosmos-3 Edge (4B) — targets board-level and GPU-less devices such as robot onboard controllers and in-vehicle units.
  • Architecture: Mixture-of-Transformers Dual Tower

    Cosmos 3 uses a Mixture of Transformers (MoT) dual-tower design. One tower handles video and image tokens; the other handles text, action, and sensor signals. The two branches exchange information via cross-attention at intermediate layers. This differs from Cosmos 1 (single Transformer) and Cosmos 2 (late-fusion).

    Training Data

    The training corpus totals roughly 1.3 billion data points (not tokens), distributed approximately as:

  • Vision: ~85 million hours of simulation video plus real-world driving and robot teleoperation clips
  • Action: ~450 million sequences (robot joint trajectories, driving trajectories, dexterous-hand actions)
  • Text / annotations: ~280 million entries (task descriptions, scene labels)
  • Sensors: ~120 million entries (IMU, torque, depth)
  • NVIDIA did not disclose parameter counts beyond the size tiers, total training tokens, or GPU-hours — consistent with the company's long-standing policy of releasing model weights while keeping training-cost figures opaque.

    Licensing

    Cosmos 3 is released under the NVIDIA Open Model License (W) v1.1 (OpenMDW). Commercial use, modification, derivative work, and redistribution are permitted, subject to:

  • No use for military or weapons systems
  • No generation of synthetic media that violates NVIDIA's Acceptable Use Policy (e.g., deepfakes)
  • Derivative models must retain the OpenMDW license
  • This is a relaxation from the non-commercial terms attached to Cosmos 1 and 2, allowing third parties to build products directly on Cosmos 3.

    Reported early adopters include 1X, Boston Dynamics (partial research collaboration), Figure AI, Skydio, an Uber ATG spinoff team, Tsinghua X-Lab, UC Berkeley BAIR, and the Stanford Fei-Fei Li group.

    Lineage and Positioning

  • Cosmos 1 (Jan 2025): established that diffusion models can generate physics-plausible video.
  • Cosmos 2 (mid-2025): added conditional control over language, action, and sensors.
  • Cosmos 3 (2026): unifies vision, action, and text within one model — eliminating the need for separate video, policy, and control stacks.
  • Relationship to the NVIDIA Stack

  • GR00T (humanoid robot foundation model): Cosmos 3 serves as its "vision brain" for scene understanding and prediction.
  • Isaac Sim / Isaac Lab: primary source of synthetic training data and the RL fine-tuning platform.
  • Omniverse: not where Cosmos 3 runs, but used as a debugging tool for inspecting generated scenes.
  • Competitive Landscape

    Compared with Google DeepMind's Genie 3, Google's Veo 3, and OpenAI's Sora 2, Cosmos 3's distinguishing feature is treating action tokens as first-class citizens. The other three models treat video as the primary output; Cosmos 3 outputs action and video on equal footing. The shift aligns with the broader industry reading that world models are moving from "generator" to "reason-then-render."

    Caveats and Open Questions

  • Training compute, total tokens, and per-tower parameter counts are undisclosed.
  • The "data point" definition (frame, segment, or token) is not clarified.
  • Public lists of the founding coalition and early adopters are incomplete; marquee names such as Boston Dynamics should be independently verified given NVIDIA's marketing tendencies.
  • No official benchmarks compare Cosmos 3 against Cosmos 1/2 or against Genie 3, Veo 3, and Sora 2.
  • Legal review of OpenMDW 1.1's clauses on training-data provenance and derivative-model weights is advised.
  • Sources

  • NVIDIA Blog (2026-08-06): https://blogs.nvidia.com/blog/open-world-models-physical-ai
  • Cosmos repository: https://github.com/NVIDIA/Cosmos
  • NVIDIA OpenMDW 1.1 license: https://developer.download.nvidia.com/licenses/nvidia-open-model-license-agreement-june-2024.pdf
  • Reuters / CNBC (2026-08-06): https://www.reuters.com/technology/nvidia-cosmos-3-physical-ai-2026-08-06/
  • VentureBeat: https://venturebeat.com/ai/nvidia-cosmos-3-physical-ai-gpt-moment/
  • The Decoder: https://the-decoder.com/nvidia-cosmos-3-world-foundation-model/
  • Key Points

  • Cosmos 3 is a three-size model family (64B / 16B / 4B) sharing a Mixture-of-Transformers dual-tower architecture.
  • Training data totals ~1.3B points across video, action, text, and sensors; compute and token totals remain undisclosed.
  • Released under OpenMDW 1.1, enabling commercial derivatives while excluding military and deepfake use.
  • Integrates with GR00T, Isaac Sim/Lab, and Omniverse as a unified physical-AI foundation.
  • Differentiates from Genie 3, Veo 3, and Sora 2 by making action a first-class output, not just video.

Tags

#nvidia#cosmos-3#world-model#physical-ai#robotics#mixture-of-transformers#autonomous-driving#foundation-model

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178603066