NVIDIA widens its free AI model lineup at GTC Berlin, from language and robots to protein design
NVIDIA released a batch of open models at GTC Berlin: new Nemotron language models for agents, Kumo models for business data, Cosmos world models for robots, a self-driving model and a protein design model built with the Baker Lab.
By ScoviAI Team
Key points
- Announced at GTC Berlin in October 2026
- New Nemotron 3.5 Super VL and Nemotron 3.5 Bolt models, built with partners including Mistral and Black Forest Labs
- Kumo: open models that forecast, score fraud risk and recommend actions from tables of business data
- Also Cosmos 3.5 world models, Alpamayo 2 Edge for cars, BioNeMo protein design and Ising for quantum computing
NVIDIA is best known for the chips that run AI. Increasingly, it also gives away the AI. At GTC Berlin, the company expanded its family of open models, with a pitch aimed squarely at Europe: models that organizations can customize with their own data, in their own languages, on infrastructure they control.
What NVIDIA released
| Family | What it's for |
|---|---|
| Nemotron 3.5 Super VL and Nemotron 3.5 Bolt | Language models for AI agents; handle text and images |
| Nemotron Retriever | Search building blocks: embedding, reranking, document parsing |
| Kumo | Predictions from structured business data |
| Cosmos 3.5 | "World models" for robots, self-driving and vision agents |
| Alpamayo 2 Edge | A compact reasoning model for autonomous vehicles |
| BioNeMo | Healthcare and scientific discovery, including protein design |
| Ising | Quantum computing |
Nemotron: models for agents
Nemotron is NVIDIA's family of open language models, datasets and training recipes for building agents. NVIDIA says they help agents run workflows, reason through problems and finish specialized tasks at lower cost. They're built with input from the Nemotron Coalition, which includes European partners Black Forest Labs, H-Company and Mistral. Companies including SAP, Salesforce, Shopify, Dassault Systèmes and WPP are using them to build custom models.
Kumo: AI for spreadsheets and databases
Most of a company's important data sits in tables: transactions, customers, inventory. Kumo models are built for that kind of data. NVIDIA says they can forecast outcomes, score fraud risk and recommend actions without building a separate machine-learning pipeline for each task. Coinbase, Databricks, DoorDash, Revolut and Tubi are among the first users, and Snowflake customers are already using them.
Robots, cars and proteins
Cosmos world models help robots and cars understand the physical world. NVIDIA added tools to train them with less compute and a Cosmos Data Factory for building training data; Porsche, Zalando and Siemens are among the users. Alpamayo 2 Edge brings scene understanding and motion planning into the car itself. In biology, BioNeMo includes a protein design model built with the Baker Lab, the group behind several major advances in protein design.
Why it matters
Open models matter for two reasons. Companies can run them privately without sending data to a US cloud, which is a big deal under Europe's data rules. And they're free to adapt, so a bank or a carmaker can tune them for its own work. For NVIDIA, it's also smart business: every open model that runs best on NVIDIA hardware sells more NVIDIA hardware.
Want to know what "open" really means? Read open-weight AI models, explained. For NVIDIA's hardware news from the same week, see the OpenShell blueprint for safe agents.
Sources
ScoviAI Team is the ScoviAI newsroom. We write every article ourselves from the original announcements and link our sources. Spotted a mistake? Tell us and we'll fix it.
ADVERTISEMENT
