# NVIDIA Details Cosmos Simulation Recipes and GPU Memory Needs By Simon Yoon Canonical URL: https://www.tokenpost.com/news/technology/47030 Published: 2026-11-02T23:20:41.000Z Updated: 2026-11-02T23:20:41.000Z Section: Technology > The guidance covers vehicle, robot and surgical simulations and lists generation memory needs from 31 GiB to 150 GiB for two Cosmos models. NVIDIA published implementation recipes for using Cosmos world models in vehicle, robot and surgical simulations, giving developers guidance on model adaptation, testing and GPU deployment. Cosmos Framework supports training with text, images, video, audio and actions. Developers can use LoRA or full supervised fine-tuning, with data formats such as JSONL and WebDataset, plus support for LeRobot and custom adapters. Compatible models include Alpamayo, GR00T, Qwen and pi0.5, alongside Cosmos 3. For autonomous-vehicle development, the Cosmos3-Nano-Transfer-Auto recipe generates multi-view video for offline testing. It supports input and output video at 30 frames per second, with clips up to 400 frames at 480p or 720p. In a reference run, NVIDIA generated seven 720p videos totaling 297 frames at 30 frames per second, along with 99 LiDAR sweeps. A separate model, Cosmos3-Nano-Sim-Auto, generates observations as a driving policy or human input changes a scenario, allowing closed-loop evaluation. The robotics recipes convert depth video into RGB scenes while preserving geometry and motion, or generate the next visual observation from a scene and robot actions. For surgical simulation, the guidance describes models that generate video from robot movements, predict actions from video or let developers test a policy as a procedure unfolds. Cosmos NIM packages supported models in optimized containers for inference. Speculative decoding delivers mean speedups of roughly two to three times on H200 GPUs. Streaming video-language-model inference can enable subsecond alerts with 10 times lower latency. The generation table lists Cosmos3-Nano at 58 GiB of GPU memory in BF16, 44 GiB in FP8 or 31 GiB with FP8 and offloading. Cosmos3-Super requires 150 GiB in BF16, 93 GiB in FP8 or 38 GiB with FP8 and offloading. Certified Cosmos NIM supports deployment on more than 20 NVIDIA GPUs. The recipes add implementation detail to earlier [Cosmos data, training and simulation tools](). NVIDIA also listed a livestream for Nov. 5. ## Links in this article - [Cosmos data, training and simulation tools](https://www.tokenpost.com/news/technology/38125)