NVIDIA released a set of open-weight Cosmos3-Nano model checkpoints on November 5, licensed under OpenMDW 1.1 for commercial and non-commercial use. It’s the actual shipping follow-through on Cosmos 3 Edge, the 4-billion-parameter world model NVIDIA previewed at its SIGGRAPH keynote back in July — the checkpoints, technical report, and training recipes are now public rather than demoed.
Four specialized checkpoints, not one general model
Rather than a single do-everything release, NVIDIA split Cosmos3-Nano into models built for specific simulation jobs:
- Cosmos3-Nano-Transfer-Auto — a bidirectional model for offline synthetic driving-sensor data, generating synchronized video across up to 11 camera views (including fisheye lenses) with optional joint RGB–LiDAR generation, guided by per-camera World Scenario Map controls and text prompts.
- Cosmos3-Nano-Sim-Auto — a causally-distilled, streaming variant of Transfer-Auto that generates successive small chunks of synchronized camera video and LiDAR in just a few denoising steps, built for real-time driving-policy simulation.
- Cosmos3-Nano-Sim-Bimanual — predicts the visual consequences of two-handed robot actions, meant as a complement to physics-simulator-driven approaches for manipulation research.
- Cosmos3-Nano-Transfer — turns a physics simulator’s own controls into generated observations, connecting simulator state directly to rendered output.
Transfer-Auto also doubles as a starting point for adapting Cosmos to a team’s own sensor rig: NVIDIA’s documentation describes post-training it on custom data to match a specific camera layout and LiDAR configuration, then optionally distilling it down into a fast streaming model the way Sim-Auto itself was built.
Why the split matters
NVIDIA frames the release around three application areas: autonomous driving, where synchronized camera-and-LiDAR generation lets a team vary environmental conditions and road layouts to expand training coverage beyond what’s available in recorded data; robotics, where Sim-Bimanual and Sim-Transfer offer complementary ways to predict what a given action will actually look like; and general world-model research, where streaming generation connects video prediction directly to ongoing interaction rather than one-shot rendering.
NVIDIA is explicit that the hard problems aren’t solved by this release — the work ahead, in its own framing, is measuring where these models faithfully support policy learning and evaluation, understanding their limits, and improving where they fall short.
Related Reading
- New NVIDIA Cosmos 3 Open Models for Interactive World Simulation — Hugging Face
- Cosmos3-Nano-Transfer-Auto checkpoint — Hugging Face
- NVIDIA’s SIGGRAPH Keynote: DLSS 5 Goes Multi-Model, and Cosmos 3 Pushes World Models to the Edge
- NVIDIA’s Technical Artists Are Directing AI Coding Agents Straight Into Houdini