Nvidia is expanding its portfolio of open models for language, robotics, biology and quantum computing. The company mainly wants to appeal to European businesses.
With the new models, Nvidia wants to provide a foundation for European companies that build applications on their own knowledge and local languages. According to the company, organisations can adapt the models to their sector. That Nvidia is making the announcement on the sidelines of the GTC conference in Berlin is no coincidence.
New language and business models
The Nemotron family includes open language models, datasets and recipes for developers building AI agents. Nemotron 3.5 Super VL processes text and images, while Nemotron Retriever focuses on embedding, reranking and document processing.
Service providers such as Accenture and EY are working with SAP, Salesforce and Shopify on custom models and domain-specific agents based on Nemotron. Nvidia is also introducing Kumo, a series of open foundation models for structured business data. These models help agents predict outcomes, assess fraud risks and recommend actions.
Robotics, vehicles and biology
Nvidia is also launching sector-specific models. Cosmos 3.5 consists of open world models that help robots, autonomous vehicles and visual agents understand the physical world. With Cosmos Data Factory, developers can compile, label and evaluate training data for driving scenarios and product defects. Porsche, Zalando and Siemens already use these tools.
Alpamayo 2 Edge is a compact reasoning model that enables real-time scene recognition and motion planning in autonomous vehicles. It runs on the DRIVE AGX Thor platform and requires less computing power. RFD4-Proteina, developed with the Baker Lab at the University of Washington, designs proteins and works about thirteen times faster than RFdiffusion3 in a benchmark with proteins of 200 amino acids.
Ising for quantum computing
The Ising models are meant to speed up the development of usable quantum computers, from hardware calibration to error correction. IQM uses Ising to calibrate its quantum processors. According to Nvidia, specialised calibration models outperform large generic models.
SEEQC combines the NVQLink platform and Ising with its own Digital Link to meet the latency requirements for real-time error correction. AWS is making Ising available to researchers through the AWS Parallel Computing Service, so they can start building before the rollout.
Nemotron 3.5 Super VL is available via Hugging Face, ModelScope, OpenRouter and directly from Nvidia. Cosmos and Ising are openly available on GitHub and Hugging Face, and Alpamayo 2 Edge follows later this month. With this expansion, Nvidia is betting on open models as the basis for sector-specific AI in Europe. The AI heavyweight is asserting itself not only on the technology front: it is also investing heavily in European start-ups.