World foundation model platform

Cosmos world model

See how world models move from generated scenes into robots, sensors, and physical decision-making.

NVIDIAPlatform launch with Cosmos 3 preview, physical-AI data-factory blueprint, and a late-May-active public GitHub stackNVIDIA platform and developer ecosystem with public GitHub repos for Predict, Transfer, Reason, Curate, Evaluator, RL, and Cookbook workflows; the repo organization remained active on May 27, 2026, while Cosmos 3 itself was still described in preview-style language and availability varies by component.
Physical AI

What this lets people do

See how world models move from generated scenes into robots, sensors, and physical decision-making.

Synthetic data, physical AI simulation, robotics, autonomous vehicle workflows, and world-model-assisted reasoning for embodied systems.

Scene explainer

Three frames before the source list.

The page starts with the experience, then moves toward source-backed details.

01

First impression

A visible world

NVIDIA Cosmos is a world foundation model platform for physical AI, with the March 2026 GTC update adding Cosmos 3 as a unifying layer for synthetic world generation, reasoning, and action simulation and the public repo surface staying active through late May 2026.

02

Capability

Why it stands out

Strong anchor for the physical AI and robotics side of world models.

03

Boundary

What not to overclaim

Cosmos should not be treated as the same category as consumer 3D world generation tools.

Good reasons to open this page

  • Visitors who want the fastest visual handle on this model lane.
  • Creators comparing whether the output feels like a clip, a place, or a controllable world.
  • Readers who need status and sources after the first impression.

Strengths

  • Strong anchor for the physical AI and robotics side of world models.
  • Useful for separating interactive creative worlds from simulation infrastructure.
  • Backed by a platform narrative rather than a single consumer demo.
  • The March 2026 update makes the product framing clearer by explicitly tying world generation, reasoning, and action simulation together under one physical-AI stack.
  • The May 2026 GitHub surface makes the stack easier to audit because curation, evaluation, and post-training tooling are visible alongside the core world foundation models.

Limits and source boundary

  • Cosmos should not be treated as the same category as consumer 3D world generation tools.
  • Hardware, SDK, and enterprise workflow context matters.
  • NVIDIA's March 16, 2026 wording mixes generally available components with preview and early-access elements, so availability should be stated precisely rather than flattened into a single public release claim.

Decision guides

Release signals

Only the selected updates that affect this profile.

The company profile stays stable. These short signals explain what changed and point back to sources.

Sources

FAQ

Dossier FAQ

Use these notes to keep model comments grounded in official sources and careful category boundaries.

Definition

What does World Models Watch count as a world model?

The site tracks systems that model environments, actions, spatial structure, or persistent simulated state. Pure text chatbots and ordinary video generators are only included when they provide a clear bridge toward interactive or physical world modeling.

Category boundary

Why do some AI video systems appear on a world-model site?

Video models are included only when they help explain the path from generated clips to controllable spaces, physics-aware prediction, or agent-ready simulation. The site keeps that distinction explicit so video generation is not overstated as a finished world simulator.

Editorial policy

How does the site decide whether a release is reliable enough to list?

Primary sources carry the most weight: official product pages, research posts, papers, documentation, code repositories, and company announcements. Secondary media can be referenced, but it stays labeled as reported or adjacent unless independently confirmed.

Community

What should readers post in comments?

Useful comments add source links, corrections, release-status notes, comparison questions, or concrete reader context. Comments are public immediately, so readers should avoid private information and unsupported promotional claims.

Read the full FAQ

Discussion

Reader discussion

Add source-backed corrections, questions, or notes for this page.

0 comments
Comments are ready in the codebase. Configure NEXT_PUBLIC_SUPABASE_URL, NEXT_PUBLIC_SUPABASE_PUBLISHABLE_KEY, SUPABASE_SECRET_KEY to enable Supabase-backed discussion in production.

No comments yet. Start with a source note or a question for future coverage.