NVIDIA's open models dominate ICML 2026 in Seoul

The International Conference on Machine Learning (ICML) 2026, running from July 6 to 11 in Seoul, South Korea, has confirmed what the research community has long been sensing: open models have become the backbone of academic AI progress, and NVIDIA sits at the very center of that shift.

Why citation count is an indicator, not a marketing figure

Citation counts in scientific papers are an objective measure of influence — researchers do not list tools and models they have not actually used. That is precisely why the figure of ~2,000 accepted papers at ICML 2026 listing NVIDIA GPUs as infrastructure speaks to how deeply that platform is embedded in the global research ecosystem. An even more concrete indicator: 145 papers directly cite Nemotron models or datasets that NVIDIA distributes under open licenses — models whose weights and training datasets are publicly available to researchers.

NVIDIA is present at the conference with 74 accepted papers across various categories, from robotics and computer vision to reinforcement learning and inference.

The open model family

NVIDIA’s open model portfolio is divided by application domain. Nemotron covers general-purpose language models; alongside the model architecture itself, NVIDIA publicly releases the datasets used in training, which is particularly valuable to researchers as it enables reproducibility and building on existing results.

The Cosmos 3 series of omni-models is designed for robotics and autonomous driving. The model understands the physical laws of the environment and can predict the consequences of actions in three-dimensional space. It is precisely Cosmos 3 that underpins DreamDojo — one of the more prominent papers of this ICML edition. DreamDojo teaches robotic behavior exclusively from videos of human activities: a robot observes how people perform a task, and Cosmos 3 helps it build an internal model of the physical environment without prior task-specific programming. Robots from Boston Dynamics, Agility Robotics, and 1X are already adopting Cosmos and Isaac GR00T models for humanoid robotics.

For the life sciences, NVIDIA is developing the BioNeMo platform. Two standout examples from ICML: the FLIP2 benchmark for predicting the effects of protein mutations, enabling researchers to standardize evaluation of new methods, and the KERMT model for predicting molecular properties in the context of drug discovery. Pharmaceutical giant Merck & Co. is already integrating KERMT into its own research processes.

Synthetic data as the dominant theme

Analysis of accepted papers shows that synthetic data generation has emerged as one of the most widely represented topics of the entire conference. The trend is not coincidental: on one side, collecting and labeling real data is becoming an increasingly costly and time-intensive process, while synthetic data generated as high-quality variations accelerates training without requiring new annotations.

KiloCode vividly illustrates the business dimension of this approach — by using Nemotron models and a code-routing architecture, the company achieved a reduction in cost per token of up to 90% compared to earlier approaches.

The broader partner ecosystem

Open models create a flywheel effect: Sakana AI builds Fugu models on the foundations of the Nemotron architecture, NAVER is developing Korean-language AI, and Together AI hosts Nemotron models as an API service for the broader developer community. In the life sciences, Basecamp Research is releasing EDEN — a DNA foundation model that relies on NVIDIA’s infrastructure.

The common thread across all these collaborations is access to open weights and datasets that lowers barriers to entry for research groups without their own resources to train from scratch.

ICML 2026 runs through July 11, and the full paper program is available to researchers through the official conference website.