NVIDIAIn use

Cosmos 3 Edge

The smallest model in the Cosmos 3 family, built on NVIDIA Nemotron to run on the robot itself. It helps robots and vision AI agents understand their surroundings, reason in real time and generate robot actions on edge computers, including Jetson Thor and the new T3000 and T2000. Weights were posted on Hugging Face on 20 July 2026.

• 4B parameters
• Open weights on Hugging Face (nvidia/Cosmos3-Edge), OpenMDW1.1 license; model card says ready for commercial use
• Tested platforms include Jetson AGX Thor and Jetson AGX Orin
• Generates video, images and actions; reasoner context up to 256K tokens

Cosmos 3 (Super and Nano)

NVIDIA's open world foundation model for physical AI, combining vision reasoning, world generation and action prediction in one model. Robot makers use it to generate synthetic training data, test robot policies and post-train robot models. Announced at GTC in March and released on 1 June 2026; robotics users named include Agile Robots, Doosan Robotics, LG Electronics, Samsung and Skild AI.

• Sizes: Cosmos 3 Super 64B, Cosmos 3 Nano 16B (Edge 4B released separately)
• Open models downloadable from Hugging Face; code on GitHub; deployable as NVIDIA NIM microservices
• Super and Nano available from 1 June 2026

Isaac GR00T N2

Next-generation GR00T robot foundation model, previewed by Jensen Huang in his GTC 2026 keynote. NVIDIA says it helps robots succeed at new tasks more than twice as often as leading vision-language-action models, and that it ranks No. 1 on the MolmoSpaces and RoboArena benchmarks. NVIDIA says it is slated to be available by the end of 2026.

• Ranked No. 1 on MolmoSpaces and RoboArena (per NVIDIA, March 2026)
• Availability: slated for the end of 2026

Isaac GR00T N1.7

NVIDIA's current open vision-language-action model for humanoids and other robots: it takes camera images, language instructions and robot state and outputs motor actions. Early access with a commercial license opened at GTC in March 2026; weights are on Hugging Face and it joined Hugging Face's LeRobot library in July. Integrators named include Techman Robot, Enactic and Nexuni.

• 3B parameters (nvidia/GR00T-N1.7-3B)
• Open weights under the NVIDIA Open Model License; model card says ready for commercial use
• Backbone: Cosmos Reason 2 vision-language model
• Pretrained on 20,000 hours of human egocentric data
• On Jetson AGX Thor: 10.7 Hz end-to-end with TensorRT
• GitHub README labels N1.7 a General Availability release (no date given)

Jetson T4000

A lower-cost Blackwell-based module in the Jetson Thor family, launched at CES 2026 as an onboard computer for running AI inside robots. NVIDIA says it delivers four times the performance of the previous generation and announced it as available now. It sits below the flagship Jetson AGX Thor T5000 in the range.

• 1,200 FP4 TFLOPS
• 64GB memory
• Configurable 70-watt power envelope
• NVIDIA Blackwell architecture

Price$1,999 at 1,000-unit volume

Isaac GR00T N1.6

Open reasoning vision-language-action model for humanoid robots and the predecessor of N1.7. Announced in September 2025 as available soon on Hugging Face, it was released there and added to Hugging Face's LeRobot library by CES in January 2026. NVIDIA said Franka Robotics, NEURA Robotics and Humanoid use GR00T-enabled workflows to simulate, train and validate robot behaviors.

• Open model on Hugging Face
• Available in the LeRobot library for fine-tuning and evaluation

Jetson AGX Thor (T5000)

NVIDIA's flagship robot computer, the onboard module that runs AI models such as GR00T directly on humanoids and other robots. Built on a Blackwell GPU, it went on sale in August 2025 as a developer kit and T5000 production module. Early adopters named by NVIDIA include Agility Robotics, Amazon Robotics, Boston Dynamics, Caterpillar, Figure and Meta.

• Up to 2,070 FP4 TFLOPS
• 128GB memory
• 130-watt power envelope
• Up to 7.5x the AI compute and 3.5x the energy efficiency of Jetson Orin

PriceDeveloper kit from $3,499