Nvidia Releases Cosmos 3 Nano Recipes for Synchronized Camera Video and LiDAR
Summary
NVIDIA released Cosmos 3 Nano recipes and checkpoints that generate synchronized camera video and LiDAR for autonomous vehicle, robot, and surgical scenarios. The company says Cosmos NIM microservices deliver roughly 2 to 3 times faster inference on H200 GPUs through quantization, speculative decoding, and KV-cache-aware routing. Cosmos3-Nano generation needs 58 GiB in BF16 and 44 GiB in FP8 with offloading.
Key Points
- NVIDIA's Transfer-Auto reference run produces seven 720p videos at 30 frames per second plus 99 LiDAR sweeps.
- Cosmos3-Nano-Sim-Auto runs autoregressive streaming generation so generated observations return to an AV policy for closed-loop evaluation.
- NVIDIA schedules a livestream on Thursday, Nov. 5 covering how developers use Cosmos Framework.