NVIDIA Unveils Cosmos 3 Nano Open Models for Interactive World Simulation

NVIDIA’s Cosmos platform expands with four new post‑trained Nano models, each carrying 16 billion trainable parameters and released under the permissive OpenMDW 1.1 license. The checkpoints target two high‑impact domains: autonomous driving and robotic manipulation.

• Cosmos3‑Nano‑Transfer‑Auto – produces up to 11 synchronized camera views plus LiDAR, with controllable weather, lighting, and traffic‑signal states.
• Cosmos3‑Nano‑Sim‑Auto – streams paired camera and LiDAR observations for closed‑loop driving simulation.
• Cosmos3‑Nano‑Sim‑Transfer – converts simulator controls and a text prompt into photorealistic video.
• Cosmos3‑Nano‑Sim‑Bimanual – predicts visual outcomes from robot or camera actions, supporting bimanual task learning.

For autonomous‑vehicle researchers, the joint RGB‑LiDAR generation ensures that appearance and geometry stay consistent, while scenario maps let teams stress‑test policies across diverse road layouts. In robotics, the action‑conditioned video prediction closes the loop between perception and control, a prerequisite for reliable sim‑to‑real transfer.

All four models share the same Nano backbone, making it straightforward to fine‑tune them on proprietary sensor rigs or domain‑specific data. The open weights and published post‑training recipes give the community a common baseline for benchmarking, ablation studies, and collaborative improvement.

Why it matters for GPU and AI infrastructure

Training and serving 16 B‑parameter world models demands high‑throughput GPU clusters with fast interconnects and large VRAM. The availability of open checkpoints lets infrastructure teams profile real‑world workloads, optimize kernel libraries, and validate scaling strategies on a representative generative‑simulation pipeline — directly informing capacity planning for AI‑GPU clouds.