Nvidia launched a family of open reasoning AI models designed for agentic AI as well as new world foundation models.
The company launched the Nvidia Llama Nemotron reasoning models that are designed for on-demand AI reasoning. Nvidia took the Llama models and enhanced them during post training to improve multistep math, coding, reasoning and complex decision-making.
According to Nvidia, the refinements made to Llama boosted accuracy by 20% compared to the base model and optimized inference speed 5x. Llama Nemotron models land with support from a variety of partners including Accenture, CrowdStrike, Microsoft, SAP and ServiceNow.
Llama Nemotron models are available as Nvidia NIM microservices in Nano, Super and Ultra sizes for various deployments.
- Nano is geared toward PCs and edge devices.
- Super is designed for the best accuracy and throughput on a single GPU.
- Ultra is designed for multi-GPU servers.
Nvidia's bet is that by open sourcing tools, datasets and post-training optimization, enterprises will build custom reasoning models.
While Llama Nemotron is focused on agentic AI, Nvidia is also pushing into physical AI models for robotics. Nvidia launched a set of new Cosmos world foundation models.
The Cosmos models include:
- Cosmos Transfer, which ingest structured video inputs such as maps, depth maps and lidar scans to create photoreal video outputs. Cosmos Transfer will streamline AI training as well as simulations and ground truth.
- Cosmos Predict, which will enable multi-frame generation and predict intermediate actions.
- Cosmos Reason, a world foundation model that will offer chain-of-thought reasoning in natural language.
For good measure, Nvidia announced Isaac GR00T N1, a humanoid robot foundation model.
Isaac GR00T N1 includes generalized skills and reasoning to human robots.
Nvidia is surrounding the model with simulation frameworks and blueprints. The company said the Isaac GR00T Blueprint for generating synthetic data as well as Newton, an open-source physics engine, will go with the new humanoid robotics model. Google DeepMind, Nvidia and Disney Research will collaborate on Newton.
Isaac GR00T N1 has a dual system architecture including a fast thinking action model and a slow thinking one that's for deliberate decisions. Key points:
- System 2 is powered by a vision language model that reasons about its environment and instructions to plan action.
- System 1 then translates system 2 data into robot movements. System 1 is trained on human demonstration data and synthetic data.
Isaac GR00T N1 can generalize tasks such as grasping and moving objects.
