Nvidia announced on 21 October an expansion of its open model portfolio across language, robotics, biology and quantum computing. The announcement brings models for searching documents and working with structured business data alongside systems for autonomous vehicles and protein design, giving organisations more ways to adapt models to their own data and run them on infrastructure they control, according to Nvidia.
Key points
- Nvidia’s expanded range includes Nemotron Retriever, Kumo, Cosmos 3.5, Alpamayo 2 Edge, BioNeMo RFD4-Proteina and Ising.
- Nemotron 3.5 Super VL handles text and visual material, while Kumo targets predictions from structured business records.
- Nvidia reports a protein-design speed result for BioNeMo; Navan reports an internal evaluation of Super VL in travel tasks.
- A smaller Bolt model is expected on 23 October.
Nemotron 3.5 meets retrieval and Kumo
The expansion follows Nvidia’s release of Nemotron 3.5 Super VL and two Bolt models, covered earlier by AI Affairs. Super VL works with text, images, documents and video. In an engineering account published on 15 October, model-hosting company Nscale described it as a 120-billion-parameter mixture-of-experts model with 12 billion active parameters. That arrangement routes work through a portion of the model’s specialised components rather than using all its parameters for each request, a design Nscale says balances capability with inference efficiency.
Checking a travel document against a policy could draw on both the page’s visual layout and its words before using the result in a multi-step request. Whether it would work for a particular set of documents and rules would depend on how the model was adapted to that task.
Nscale also describes changes aimed at long-running agents, which repeatedly process new information and use tools: the model’s architecture and multi-token prediction are intended to speed generation, while compression of the memory used to retain earlier context gives more room for simultaneous sessions. In Nvidia’s agentic coding benchmark, Super VL completed 10,000 tasks up to 4× faster than Qwen3.8 Flash Next on 4× Nvidia Blackwell GPUs, according to Nscale’s account of Nvidia’s comparison.
Nvidia says its Nemotron models draw on data and research from the Nemotron Coalition, which includes Black Forest Labs, H-Company and Mistral. The newly announced Nemotron Retriever models cover embeddings, reranking and document parsing — ways to represent information for search, reorder results and extract material from documents. Nvidia’s Kumo range instead addresses records arranged in tables and related datasets, with models for prediction, forecasting and anomaly detection and an open library for developing and deploying them. Nvidia says organisations can fine-tune Kumo models on their own data.
Cosmos 3.5 and Alpamayo 2 Edge
For systems that must deal with physical surroundings, Nvidia lists Cosmos 3.5 world models and Alpamayo 2 Edge among the expanded portfolio. It says Cosmos helps developers build robots, autonomous vehicles and vision agents, and is introducing checkpoints for testing behaviour in simulated driving and robotic-manipulation scenarios before real-world tests. Its new Cosmos Data Factory provides tools for preparing, labelling, generating and assessing training data, including driving scenarios and product defects.
Alpamayo 2 Edge takes a different part of the driving problem into the vehicle. Nvidia describes it as a compact model for understanding a scene and planning motion in real time, built on a Cosmos 3 Edge foundation based on Nemotron and optimised for its DRIVE AGX Thor platform. That places the computation where the vehicle runs, while Cosmos also supplies simulated situations for development and evaluation.
Nvidia says Porsche with Voxel51, ANYbotics with Encord, Zalando, CEVA Logistics with Sereact, and Siemens are using Cosmos Data Factory to build training datasets and improve model precision. Its account names them as users of the data tools, rather than as deployments of Alpamayo 2 Edge.
BioNeMo protein designs and Navan’s Super VL test
BioNeMo RFD4-Proteina, developed by Nvidia with the Baker Lab at the University of Washington’s Institute for Protein Design, generates protein sequences and three-dimensional structures. Nvidia says it produced designs about 13x faster than RFdiffusion3 in a benchmark involving 200 amino acid proteins. Researchers can use the model to generate and assess candidate designs against constraints before selecting designs for laboratory tests, Nvidia says. The company also includes an Ising family for quantum computing in the expanded range.
Among organisations working with Nemotron models and datasets, Nvidia names Dassault Systèmes, Salesforce, SAP, Shopify, Synopsys and WPP, as well as Accenture and EY. A more detailed account of one deployment comes from travel and expense platform Navan: in a statement issued on 16 October, it said it had trained and deployed Super VL within three weeks of receiving early access, and that the model was powering several of its travel agents in production.
Navan says it tested a fine-tuned version against its existing self-hosted model and other models using consistent prompts, tools, travel scenarios and criteria. In that internal evaluation, Super VL beat or tied competing models on six of seven measures and had the lowest latency, according to Navan. The test concerns its own travel tasks and its fine-tuned deployment, a more specific setting than Nvidia’s coding benchmark.
The two Bolt models serve smaller deployments. Nemotron 3.5 Bolt 74B-A9B 24GB is available alongside Super VL, Verdict reported in an account carried by Yahoo Finance. Nvidia expects to launch Nemotron 3.5 Bolt E4B 4GB on 23 October, Verdict reported.
