NVIDIA details Cosmos Framework: world model training, sim recipes, and NIM deployment
NVIDIAAI · x · 2026-11-03
NVIDIA published a technical blog on Cosmos, its world foundation model platform for physical AI.
Key points
- Cosmos Framework: modular open tools for training, RL, optimization, and inference across Cosmos 3, Alpamayo, GR00T, Qwen, and pi0.5
- Cosmos 3 Nano sim recipes: released checkpoints generate synchronized camera video + LiDAR for AV, robotics, and surgical scenarios; depth video converts to RGB with new lighting/textures while preserving geometry and motion
- Cosmos NIM microservices: quantization, speculative decoding, and KV-cache-aware routing deliver 2-3x speedups on H200 GPUs
- GPU memory: Nano generation needs 58 GiB (BF16) or 44 GiB (FP8, 31 GiB with offloading); Nano reasoning 23 GiB
A livestream on Nov 5 walks through the framework.
Related event: NVIDIA Releases Cosmos Simulation Recipes and World Model Framework(4 posts)→
More from Embodied
- Industrial designer uses fal to iterate real hardware designs from photos, faster prototyping — OdinLovis · 2026-11-03
- NVIDIA Cosmos Simulation Recipes Turn Daytime Drives Into Rainy Nights With Trajectories Intact — NVIDIAAI · 2026-11-03
- AI Agent Performs Brain Drug Injection on a Live Mouse — ChuckDBrooks · 2026-11-03
- 3D World Simulator predicts appearance, geometry and touch at 50+ FPS on an RTX 5090 — YunzhuLiYZ · 2026-11-03
- Google's Suncatcher satellite runs Gemma 270mi n space—power limits mean only one TPU lit at a time — demian_ai · 2026-11-03
- Apple targets 2027 launch for $1,000 home robot, code-named J595 — Ars Technica AI · 2026-11-03