NVIDIA Demos 1-Day Post-Training for Cosmos 3

NVIDIA Developer · youtube · 2026-07-17

NVIDIA explained how to complete the post-training of Cosmos 3 in a single day using **TAO Agentic Skills**, streamlining the workflow with less manual effort. Key takeaways include: - The goal is to solve the "last mile" accuracy problem in physical AI, specifically addressing data bottlenecks and slow post-training iterations. - The livestream demonstrated triggering an end-to-end post-training pipeline from a single natural language prompt, covering **LoRA vs. SFT** selection, AutoML hyperparameter tuning, and post-training deployment via **NIM**. - Official benchmark results showed that the video QA accuracy of Cosmos 3 Nano improved from **54.41% to 93.35%**.

Related event: NVIDIA Demonstrates One-Day Post-Training for Cosmos 3(2 posts)→

Original post →

More from Infra

Infra channel →