What's New in NVIDIA JetPack 7.2?

With the release of JetPack 7.2 SDK, NVIDIA expands the JetPack 7 software stack beyond Jetson Thor by adding support for the Jetson Orin family. This brings Orin and Thor onto a common software foundation built on Ubuntu 24.04, Linux Kernel 6.8, and CUDA 13.

JetPack 7.2 also introduces several new capabilities, including:

• Official Yocto support
• AGX Orin 32GB Super Mode
• NVIDIA agent skills
• NemoClaw integration
• Multi-Instance GPU (MIG) support on Jetson Thor

In this blog, we highlight the most relevant JetPack 7.2 updates for rugged edge AI systems and how developers can take advantage of them in production deployments.

Source: NVIDIA

JetPack 6.2 vs JetPack 7.2: What Changed?

Before we get into details, here's a high-level look at the key updates from JetPack 6.2 to JetPack 7.2 for Jetson developers.

For existing Jetson Orin users, JetPack 7.2 brings the platform onto the same JetPack 7 software foundation as Jetson Thor while providing updated AI libraries, development tools, and capabilities for increasingly complex edge AI workloads.

1. Official Yocto Support for Jetson 

One of the most significant additions in JetPack 7.2 is official support for the Yocto Project, making it easier for developers to build custom Linux distributions for Jetson-based systems.

For many developers, Ubuntu-based JetPack images provide the fastest path to getting started with Jetson development. However, as applications become more specialized, a general-purpose Linux distribution may include software packages, services, and dependencies that are unnecessary for the target workload.

Yocto takes a different approach by allowing developers to build a Linux image from the ground up, selecting only the components required for their application.

What Is Yocto?

The Yocto Project is an open-source framework for creating custom Linux distributions for embedded systems.

Instead of deploying a full Ubuntu image, developers can build a streamlined operating system containing only the packages and services required by their application.

For example, a computer vision appliance may only need Linux kernel, CUDA, TensorRT, DeepStream, and Application software, while excluding desktop environments, Bluetooth services, printing services, and other unused components.

There are several advantages of having a customized Linux image:

• Smaller software footprint
• Lower memory usage
• Reduced attack surface
• Faster boot times
• More predictable system behavior
• Greater control over software updates

2. AGX Orin 32GB Super Mode

As JetPack 7.2 extends support to the Jetson Orin family, NVIDIA also introduces Super Mode for Jetson AGX Orin 32GB, increasing AI performance from 200 TOPS to 241 TOPS.

This gives developers additional performance headroom on the AGX Orin 32GB platform without moving to the higher-cost AGX Orin 64GB module. For compute-bound workloads, Super Mode can help improve inference throughput or support more demanding model pipelines while keeping the system based on a lower-cost Orin configuration.

What Changed?

With Super Mode enabled, Jetson AGX Orin 32GB can operate at a higher performance profile. According to NVIDIA, GPU frequency increases from 930 MHz to 1.3 GHz, with the power envelope increasing up to 60W.

Source: NVIDIA | Jetson AGX Orin module comparison

Super Mode gives developers a middle ground between AGX Orin 32GB standard operation and AGX Orin 64GB. If the workload fits within 32GB of memory but needs more compute performance, Super Mode can provide additional throughput without requiring a move to the 64GB module.

3. Multi-Instance GPU (MIG) Support on Jetson Thor

JetPack 7.2 adds Multi-Instance GPU (MIG) support on Jetson Thor, enabling a single GPU to be partitioned into multiple instances that can be assigned to different workloads.

Source: NVIDIA | How Multi-Instance GPU Works

Jetson Thor supports two MIG partitions:

This helps reduce resource contention when multiple applications are running on the same system.

By allocating dedicated GPU resources to latency-sensitive functions such as perception, sensor fusion, motion planning, and safety monitoring, developers can run AI inference or generative AI workloads on a separate partition without impacting critical workloads.

This capability is particularly relevant for robotics, transportation, smart infrastructure, and industrial automation applications that need to consolidate multiple workloads on a single edge platform while maintaining predictable performance.

4. NVIDIA Agent Skills for Jetson

JetPack 7.2 introduces NVIDIA agent skills for Jetson, a set of AI-assisted workflows designed to simplify common development and optimization tasks.

These include two types of skills:

• Jetson Device Skills: Run directly on Jetson to support system diagnostics, memory optimization, model benchmarking, inference tuning, and deployment configuration. These skills help developers analyze system resources and optimize AI workloads for the target Jetson platform.
• Jetson BSP Skills: Run from a host system to streamline BSP customization and board bring-up, including device trees, I/O configuration, power profiles, and other Jetson Linux settings. This can simplify hardware enablement when moving from development kits to custom or production carrier boards.

These NVIDIA Agent Skills address key workflows such as Jetson Linux customization, memory optimization, and model benchmarking, helping developers reduce manual configuration and accelerate system bring-up, performance tuning, validation, and production deployment.

5. NemoClaw Integration for Agentic AI

JetPack 7.2 adds support for one-command NemoClaw deployment on Jetson, making it easier to build and deploy AI agents at the edge.

NemoClaw is NVIDIA's open-source framework for orchestrating multiple AI components, including vision models, language models, local data sources, and tool execution.

Source: NVIDIA GTC 2026 Keynote

For example, an edge system could use a vision model to detect an issue, retrieve relevant information from a local database, and generate an alert or report based on the results.

NemoClaw is particularly useful for applications that require AI systems to perform multi-step reasoning and actions locally rather than relying solely on single-model inference.

To install NemoClaw on a Jetson device running JetPack 7.2, run the following single command:

curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash

Conclusion

While JetPack 7.2 introduces a wide range of new features, many of the updates point toward a common goal: enabling more complex AI workloads on Jetson platforms.

Features such as NemoClaw, NVIDIA agent skills, and memory optimization workflows suggest a growing focus on agentic AI and physical AI applications, where multiple models, tools, and services may need to run together on a single edge device. At the same time, capabilities such as Yocto support, AGX Orin 32GB Super Mode, and MIG on Jetson Thor help provide the flexibility, performance, and deployment options needed to support these increasingly sophisticated workloads.

As AI applications continue to evolve beyond single-model inference, JetPack 7.2 provides a stronger software foundation for building and deploying next-generation edge AI systems.

This blog is based on NVIDIA's technical blog, Deploy Agentic-Ready AI at the Edge with Memory Efficiency in NVIDIA JetPack 7.2.

JetPack 7.2 on Premio's NVIDIA Jetson Solutions

We're currently evaluating JetPack 7.2 across our NVIDIA Jetson-based JCO Series and WCO Series edge AI computers. We're also validating AGX Orin 32GB Super Mode on the JCO-6000-ORN Series to evaluate its impact on performance, power, and thermal behavior under real-world edge AI workloads.

Stay tuned for more updates as we continue exploring JetPack 7.2 and its impact on our NVIDIA Jetson-based edge AI computers.

For any inquiries, feel free to contact us here.

This blog was originally published on June 24, 2026, and has been updated to reflect the latest product and certification information.