NVIDIA AI Updates: July 15, 2026

1. NVIDIA Shows One-Day Post-Training of Cosmos 3 Physical AI Model With TAO Agent Skills

NVIDIA. NVIDIA published a technical walkthrough showing how to post-train its open physical AI foundation model Cosmos 3 in about a day using TAO agent skills, which let a coding agent automate dataset processing, baseline evaluation, LoRA configuration, and hyperparameter optimization. Cosmos 3 uses a mixture-of-transformers architecture unifying text, image, video, audio, and action inputs and ships in Nano (16B) and Super (64B) sizes. In the demonstration, LoRA fine-tuning raised accuracy on the Woven Traffic Safety dataset from a 54.41% zero-shot baseline to 87.14% in roughly 30 minutes, and TAO AutoML sweeps pushed it to 93.35%. Source

2. NVIDIA Details Autonomous ML Research Workflow Using RL Agent Skills and NeMo

NVIDIA. NVIDIA detailed how to run an autonomous machine learning research workflow using RL agent skills with NVIDIA NeMo RL and NeMo Gym, where reusable agent skills named Brev-etiquette, Session-memory, and Autoresearch let a coding agent handle GPU setup, experiment iteration, and paper-to-code implementation. In one demonstration the agent lifted a vision-language model’s accuracy on a custom star-counting task from 25.0% to 96.9% via supervised fine-tuning, and it reproduced the OAPL off-policy RL algorithm directly from a research paper. The workflow runs on NVIDIA Brev GPU instances while researchers retain oversight of goals and data. Source