In brief
- Nemotron 3.5 Super VL is among the new open models that support text and images.
- NVIDIA Kumo is designed for forecasting, anomaly detection and prediction using structured business data.
- NVIDIA Cosmos models and data preparation tools target training workflows for robotics and driving scenarios.
- Nemotron 3.5 Super VL is available on Hugging Face, ModelScope, OpenRouter and NVIDIA’s platforms.
On October 21, 2026, NVIDIA announced its open model portfolio for organizations and developers in Europe in a company blog post. The new announcement covers models spanning language and image processing, structured business data, robotics, biology and quantum computing. NVIDIA says organizations can customize the models with their own data and use them on infrastructure they control. The newly announced products include Nemotron 3.5 Super VL, Nemotron Retriever, Kumo and Cosmos models.
What tasks does the Nemotron family target?
NVIDIA describes Nemotron as a foundation for building agents that perform specialized tasks, combining open language models, datasets and customization recipes. The company says the models support text and images and can be used for tasks such as managing workflows and reasoning through complex problems. Nemotron 3.5 Super VL is one of the new models in the announcement.
Nemotron 3.5 Super VL is available through Hugging Face, ModelScope and OpenRouter, as well as through NVIDIA’s cloud and partner services. It is also offered as an NVIDIA NIM microservice on build.nvidia.com. Nemotron Retriever models are designed for embedding, reranking and document parsing. The expected availability date announced for Nemotron 3.5 Bolt E4B 4GB is October 23.
The scope of Kumo and Cosmos
NVIDIA Kumo focuses on tasks such as forecasting, prediction and anomaly detection with tabular and relational data. The portfolio includes Kumo Structured, which incorporates these models, and the open Kumo Engine library stack, a set of tools for developing and deploying structured-data models. According to NVIDIA, ready-made models can be added to applications, while fine-tuning with custom data is an option when higher performance is needed. Kumo models are available on Hugging Face, and an NIM microservice packaged for production inference is available on NVIDIA’s platform.
Cosmos is presented as a family of world models for understanding and interacting with the physical world. NVIDIA’s new post-training refinement and optimization tools are intended to improve models while using less compute. Cosmos Data Factory covers workflows for preparing, labeling, generating and evaluating training data in areas such as driving scenarios and product defects. Cosmos models are available through Hugging Face and GitHub.
Potential implications for architecture and visualization teams
The announcement does not describe a feature built directly into architectural modeling or rendering software. However, Nemotron’s support for visual and text inputs, along with Retriever models focused on document parsing, offers areas for teams to explore when assessing workflows involving documents and images. The source does not say that these models will automate any specific process in architecture firms.
Kumo’s support for tabular and relational data may interest teams working with structured project data, but no architectural use case or compatible software is specified. Before testing, firms should separately check model access terms, options for customization with their own data, and the requirements of the deployment infrastructure. NVIDIA says the models can be used on infrastructure controlled by organizations; this announcement does not provide hardware requirements or licensing details.
Availability and details not covered in the announcement
The announcement lists access channels for models such as Nemotron Retriever, Kumo and BioNeMo RFD4-Proteina. NVIDIA also says it has made Cosmos models available on GitHub and Hugging Face. Alpamayo 2 Edge is expected to be available for download through GitHub and Hugging Face toward the end of the month. This timeline is presented as an expectation in the announcement.
The announcement brings together products developed for different industries and tasks rather than focusing on a single model. However, each model requires separate information on technical requirements, commercial use terms and compatibility with architecture software. Teams should therefore consult product-specific access and usage documentation when assessing how the portfolio fits their workflows.
Sources
1 sourceSource texts are not republished; short quotes are marked, everything else is our own summary and commentary.
For architecture firms and visualization teams, the announcement’s practical value lies not in offering a ready-made rendering or BIM feature, but in bringing together a range of components for developing agents focused on text, images and documents. Document parsing and structured-data models can provide a meaningful workflow advantage only when tested with real project data.
Teams in Turkey should clarify licensing and commercial-use terms, whether the models can run on their own infrastructure, and hardware requirements before testing. These details are not covered in the announcement, and direct compatibility with architecture software is not specified either. For now, the best approach is to monitor the products and conduct a controlled pilot evaluation.
Frequently asked questions
Where can I use NVIDIA Nemotron 3.5 Super VL?
The model is available on Hugging Face, ModelScope and OpenRouter. It is also offered as an NVIDIA NIM microservice on build.nvidia.com.
What is NVIDIA Kumo used for?
Kumo was developed for tasks such as forecasting, prediction and anomaly detection using structured and tabular data.
Where are NVIDIA Cosmos models available?
NVIDIA announced that Cosmos models are available on Hugging Face and GitHub.
