Open models take center stage at ICML 2026

Open frontier models and open AI infrastructure emerged as central themes at ICML 2026, with NVIDIA reporting 74 accepted papers. Approximately 2,000 accepted papers cite NVIDIA GPUs, while 145 cite NVIDIA Nemotron, an open model family that includes open datasets. Researchers also drew on NVIDIA Cosmos, Isaac GR00T, BioNeMo and related model families across physical AI, robotics, autonomous vehicles and biomedical work.

Research themes included vision and video generation, reinforcement learning for large language models, agent training and AI inference. Robot world models drew attention through projects such as DreamDojo, which uses NVIDIA Cosmos open frontier models to learn physical behavior from human video and evaluate robot policies in virtual environments. In life sciences, BioNeMo supported work on protein function, molecular behavior and genetic code, including FLIP2 benchmarks and KERMT for molecular property prediction.

NVIDIA framed Nemotron as a broader research stack with open weights, datasets and recipes for reasoning, tool use, safety, data curation and efficient inference. The Cosmos 3 family supports robots, autonomous vehicles and vision AI, while partners are building on the stack: Basecamp Research developed EDEN, Merck & Co. uses KERMT, Sakana AI built Fugu models on Nemotron 3 Ultra, and KiloCode reported token cost reductions of up to 90% using Nemotron in code routing.