The Pulse
NVIDIA Expands Open Models for Robots and Autonomous Cars
NVIDIA announced new open models and tools for robotics, autonomous vehicles and other fields on October 21. Its physical AI releases include Cosmos 3.5 world models, simulation tools and Alpamayo 2 Edge, a compact model for in-vehicle scen
AI.info Team · Written with AI · fact-checked ·
NVIDIA says its new protein-design model generates candidates about 13 times faster than RFdiffusion3 in a benchmark involving 200-amino-acid proteins. The same October 21 announcement puts physical AI among a broader expansion of open models, with new tools for training and testing robots and autonomous vehicles before they operate in the real world.
Cosmos adds simulation before real-world tests
The company’s Cosmos family is aimed at developers building robots, autonomous vehicles and vision agents that interact with physical environments. NVIDIA says its new post-training and optimization tools are designed to improve world models using less compute, while simulation checkpoints let developers evaluate behavior in interactive driving and robotic-manipulation scenarios before real-world testing.
NVIDIA is also introducing Cosmos Data Factory, a set of blueprints and agent skills for curating, labeling, generating and evaluating training data. The workflows cover driving scenarios, product defects, image augmentation and event generation. Developers can run a complete workflow or use individual capabilities, with agents coordinating the work.
The company names Porsche with Voxel51, ANYbotics with Encord, Zalando and CEVA Logistics with Sereact, and Siemens as users of Cosmos Data Factory. NVIDIA says they are using it to create training datasets and improve model precision; the announcement does not provide comparative performance results for those deployments.
Alpamayo 2 Edge targets in-vehicle reasoning
For autonomous vehicles, NVIDIA introduced Alpamayo 2 Edge, a compact reasoning model intended to handle scene understanding and motion planning directly inside a vehicle. The model is built on Cosmos and optimized for NVIDIA’s DRIVE AGX Thor platform. NVIDIA says the design aims to bring advanced driving capabilities to vehicles with lower computing requirements.
The model was not immediately downloadable at announcement time. NVIDIA said it expected to make Alpamayo 2 Edge available on GitHub and Hugging Face later in October, leaving the announced release window short but not giving a specific date.
Seven model lines span industry and research
The physical-AI releases sit within a portfolio NVIDIA describes across seven areas: Nemotron language and vision models, Nemotron Retriever, Kumo models for structured business data, Cosmos world models, Alpamayo for autonomous vehicles, BioNeMo for biology, and Ising for quantum computing. The company says organizations can customize its open models with local data and deploy them on infrastructure they control.
That breadth includes products with different jobs and audiences. Kumo models are designed for tasks such as forecasting, fraud-risk scoring and recommendations from structured data, while BioNeMo RFD4-Proteina generates protein sequences and three-dimensional structures. In the protein benchmark NVIDIA cites, the speed comparison is specifically against RFdiffusion3 and concerns 200-amino-acid proteins; it is not a result for Cosmos or the physical-AI models.
Availability varies across the portfolio
NVIDIA says Cosmos is available through Hugging Face, GitHub and selected inference partners. Nemotron Retriever models and Kumo models are also available through Hugging Face, with Kumo offered as an NVIDIA NIM microservice; BioNeMo RFD4-Proteina and Ising are available through several NVIDIA and developer platforms.
Nemotron 3.5 Super VL is available through Hugging Face, ModelScope, OpenRouter and NVIDIA’s build platform, among other channels. NVIDIA said Nemotron 3.5 Bolt E4B 4GB was expected on Friday, October 23. The staggered release dates mean the announcement describes a collection of models at different stages of access, rather than a single package available all at once.