Nvidia widens its open-model lineup from tabular data to protein design

Nvidia announced a sweep of open models aimed at European developers and industries: Nemotron 3.5 Super VL and Nemotron 3.5 Bolt, new Nemotron Retriever models for search and document parsing, a Kumo family for structured business data, Cosmos 3.5 world models, Alpamayo 2 Edge for autonomous vehicles, the BioNeMo RFD4-Proteina protein design model and Ising models for quantum computing. Nemotron models draw on the Nemotron Coalition, which includes European partners Black Forest Labs, H-Company and Mistral, and Nvidia lists Dassault Systèmes, SAP, Salesforce, Shopify, Synopsys and WPP among companies building custom models on Nemotron. Bolt was expected Friday, October 23 (Nemotron 3.5 Bolt), and Alpamayo 2 Edge later in October.

The pitch is that organizations take these as a foundation, tune them on local data and run them on infrastructure they control, which addresses European concern about dependence on a few foreign AI providers, at least at the model layer. Sector showcases followed for banks (banks on open weights) and governments (public sector).

Kumo may be the most practically important release. Most companies need better predictions from tables they already have, such as fraud scores, forecasts and recommendations, more than a chatbot that reasons harder. Nvidia says Kumo’s pretrained models can be dropped into applications and fine-tuned on proprietary data, avoiding a separate machine-learning pipeline for each task (Kumo Relational). Coinbase, Databricks, DoorDash, Revolut and Tubi are named early users; production case studies will show whether it works as described.

BioNeMo RFD4-Proteina, built with the Baker Lab at the University of Washington’s Institute for Protein Design, is described as leading on more than 10 protein design tasks and generating designs about 13 times faster than RFdiffusion3 on a 200-amino-acid benchmark. Faster candidate generation widens the funnel for labs but does not produce therapies by itself, and the benchmark claims are Nvidia’s own; the Baker lab’s separate enzyme work showed design still needs lab validation (AI-designed proteases).

Open weights lower the cost of building, but nearly every availability note routes through Nvidia’s NIM packaging, its cloud partners and its hardware targets, such as the DRIVE AGX Thor platform for Alpamayo. The approach is open in the sense that matters for customization and also makes Nvidia the default substrate for those who customize. Licensing terms for the new weights will matter too.

Autonomous driving models and protein design tools are dual-use or safety-critical. A compact reasoning model for vehicles is welcome, but deployment should be gated by testing, which is why new Cosmos simulation checkpoints for evaluating driving and manipulation behavior before real-world trials (Cosmos tools) are as important as the models. More in our Nvidia coverage, robotics coverage and research coverage.