NVIDIA Unveils Suite of Open Models, Including Cosmos for Physical AI and Alpamayo for Autonomous Driving

On January 5, 2026, NVIDIA released new open models, data and tools aimed at advancing AI across industries. The models include the Nemotron family for agentic AI, Cosmos for physical AI, Alpamayo for autonomous driving development, Isaac GR00T for robotics and Clara for biomedicine. NVIDIA also made available one of the world's largest open multimodal datasets, including 10 trillion language training tokens, 500,000 robot trajectories, 455,000 protein structures and 100 terabytes of vehicle sensor data. Companies including Bosch, CodeRabbit, CrowdStrike, Cohesity, Fortinet, Franka Robotics, Humanoid, Palantir, Salesforce, ServiceNow, Hitachi and Uber are adopting or building with the offerings.

NVIDIA's release of open models and large-scale data in one package could significantly reduce development costs for physical AI and autonomous driving and accelerate the deployment of robots and self-driving vehicles. It could speed up AI-driven industrial transformation and have a major impact on the competitive landscape.

NVIDIA says Nemotron Speech, part of the Nemotron family and including speech-recognition models, is 10 times faster than other models in its class on the Daily and Modal benchmarks. Nemotron RAG enhances multilingual, multimodal data retrieval, while Nemotron Safety includes Llama Nemotron Content Safety models and PII detection models. Cadence and IBM are piloting Nemotron RAG, while CrowdStrike, Cohesity and Fortinet are adopting Nemotron Safety. Palantir is integrating Nemotron into its Ontology framework, and CodeRabbit is using it for AI code reviews. For physical AI, NVIDIA released Cosmos Reason 2, a reasoning VLM, and Cosmos Transfer 2.5 and Cosmos Predict 2.5 for synthetic video generation. Isaac GR00T N1.6 is an open reasoning VLA model for humanoid robots that enables whole-body control. For autonomous driving, NVIDIA released Alpamayo 1, its first open, large-scale reasoning VLA model, and AlpaSim, a simulation framework for closed-loop training. It also made available a physical AI open dataset containing more than 1,700 hours of driving data. For health care, NVIDIA announced La-Proteina for protein design, ReaSyn v2 for synthesis-aware design, KERMT for safety testing and RNAPro for 3D RNA structure prediction. It also made available a dataset of 455,000 synthetic protein structures.

For Japanese companies, NVIDIA's open models and data offer an opportunity to accelerate robotics and autonomous driving development. Isaac GR00T for humanoids and Alpamayo for autonomous driving are especially relevant to sectors in which Japanese manufacturing and automotive companies have strengths. Using these models could shorten development timelines and reduce costs. However, greater reliance on NVIDIA's platform could standardize technologies and make it harder for Japanese companies to differentiate their proprietary technologies.