NVIDIA Jetson Modules Span 67 TOPS to 2070 FP4 Teraflops for Edge Robotics

NVIDIA published a series of examples on July 28, 2026 to walk developers through its Jetson platform for edge AI and robotics, covering three modules with different performance tiers. The company framed the lineup around portability, with venture capitalist Sarah Guo demonstrating the devices in a handbag, but the concrete specs and project list give inference engineers a practical map of what each board can handle.

From 67 TOPS to 2070 FP4 Teraflops

The entry point is the Jetson Orin Nano Super developer kit, rated at 67 trillion operations per second (TOPS). NVIDIA positions it for first-time builders learning computer vision, prototyping edge AI, and building AI agents. The company also introduced Jetson Device Skills and Jetson BSP Skills, which are meant to make it easier for developers to use coding AI agents to create, optimize, and deploy edge AI workloads.

Example projects on the Orin Nano Super include a custom SidewalkPilot AI model that autonomously drives a toy electric vehicle, and the Reachy Mini Jetson Assistant, a fully on-device voice and vision assistant for the Reachy Mini Lite robot. That assistant runs locally with GPU acceleration, with no cloud, API keys, or internet required at runtime. Another example shows a first-time developer building an AI-powered robot from scratch using the open-weight Mistral model. A separate project uses the Orin Nano Super to run a Yocto-powered video podcast where two AI models discuss topics in real time.

For more complex workloads, NVIDIA lists the Jetson AGX Orin at 275 TOPS. The company targets this module at advanced coursework, capstone projects, and applied research, supporting computer vision, generative AI, and autonomous navigation. Example deployments include a browser-based interface that streams live vision language model inference through a camera feed, and SMoRes, a Carnegie Mellon University robotics team's autonomous system that builds a 3D map of an environment while searching for survivors in rescue scenarios.

At the top of the lineup, the Jetson AGX Thor delivers up to 2070 FP4 teraflops of AI compute with 128GB of memory. NVIDIA describes it as the platform for next-generation humanoid robots and autonomous systems that need real-time reasoning at the edge. Example projects include Matcha Bot, a first-place hackathon entry from University of Illinois Urbana-Champaign's SIGRobotics team that uses two robotic arms running the NVIDIA Isaac GR00T N1.5 model to autonomously pour, prepare, and whisk matcha. Another example is a Multimodal AI Studio, a development environment on the AGX Thor that combines vision, audio, and language inputs for multimodal AI experimentation at the edge.

What This Means for Model Serving

The three tiers give builders a clear scaling path. The Orin Nano Super suits small models and on-device assistants where latency and privacy matter, with no cloud dependency at runtime. The AGX Orin handles larger vision language models and navigation stacks. The AGX Thor, with 128GB of memory and FP4 compute, targets the largest workloads, including humanoid robot control with models like Isaac GR00T N1.5.

NVIDIA also announced a livestream series across three modules where developers can learn to run generative AI, build claw agents, and bring vision-language and vision-language-action models to physical AI applications on Jetson hardware. The company said it will publish more examples throughout the week, starting with the Orin Nano Super.