Edge AI (2026) | Jetson, LiteRT, ONNX Runtime, OpenVINO, Edge Impulse & More

Edge AI Hardware & Modules

8 entries

Accelerators, modules and vision hardware that run models on the device.

01NVIDIA JetsonEmbedded modules and developer kits that put GPU-accelerated AI in robots, cameras and other edge devices, with the JetPack software stack for deploying vision and generative AI models. 02NVIDIA IGXIndustrial-grade edge AI platform from NVIDIA. 03Google CoralOpen-source platform with Edge TPU accelerators and tools for building fast, private on-device inference into products. 04Axelera AIAI processing units aimed at efficient edge inference. 05SiMa.aiPhysical AI platform for embedded edge systems. 06SyntiantLow-power AI processors for always-on devices. 07DEEPXEdge AI chips for physical AI. 08LuxonisAI depth cameras and edge vision hardware.

Development & Deployment Platforms

7 entries

Collect data, train, optimise and deploy models for devices.

09Edge ImpulseDevelopment platform for collecting data, training and deploying ML models to microcontrollers, sensors and gateways, now with MLOps workflows for fleets of edge devices. Qualcomm announced in March 2025 that it would acquire Edge Impulse, and the company's website lists its address as c/o Qualcomm. 10Qualcomm AI HubHosted library of optimized models for on-device AI, where you can also compile your own model and benchmark it on real Qualcomm-powered phones, IoT and compute devices. 11Google AI EdgeGoogle's tools for deploying AI on mobile, web and embedded devices. 12RoboflowPlatform for building and deploying computer vision models. 13Latent AIEdge AI platform for optimising and securing models. 14Nota AIModel compression and optimisation for on-device AI. 15Ultralytics YOLOYOLO models for object detection and segmentation, with export to edge formats.

Inference Runtimes & Compilers

13 entries

Run optimised models efficiently on phones, gateways and microcontrollers.

16LiteRT (formerly TensorFlow Lite)Google’s runtime for running ML and generative AI models on-device, with CPU, GPU and NPU acceleration on Android, iOS, Linux and microcontrollers. 17ONNX RuntimeCross-platform inference engine that runs models exported from PyTorch, TensorFlow and others on servers, PCs, mobile and edge hardware through pluggable hardware accelerators. 18ExecuTorchPyTorch’s on-device runtime for deploying models to mobile, wearable and embedded devices without leaving the PyTorch workflow. 19Intel OpenVINOOpen-source toolkit for optimizing and deploying vision, language and generative AI models across Intel CPUs, GPUs and NPUs, from edge boxes to the data center. 20Apache TVMOpen-source machine learning compiler framework that optimizes models for CPUs, GPUs and specialized accelerators, a common path to unusual edge hardware. 21MediaPipeGoogle AI Edge toolkit of ready-made, customizable on-device solutions for vision, text, audio and generative AI tasks. 22NVIDIA TensorRTSDK for high-performance deep learning inference on NVIDIA hardware. 23NVIDIA DeepStream SDKToolkit for building streaming video analytics pipelines. 24Core MLApple's framework for running models on iPhone, iPad and Mac. 25Qualcomm AI Engine Direct SDKQualcomm's SDK for running AI on its processors. 26TensorFlow Lite MicroInference infrastructure for microcontrollers and other small devices. 27CMSIS-NNOptimised neural network kernels for Arm Cortex-M processors. 28ONNXOpen format for exchanging models between frameworks.

On-Device & Local LLMs

5 entries

Running language models locally on laptops, phones and edge servers.

29llama.cppLLM inference in C/C++ on a wide range of hardware. 30OllamaRun language models locally. 31MLC LLMUniversal LLM deployment engine using machine learning compilation. 32llamafileDistribute and run LLMs as a single file. 33LM StudioDesktop app for running local language models.

Community & Reading

2 entries

A community and a related book.

34tinyML FoundationCommunity for machine learning on ultra-low-power devices. 35A Book on AI for Smart and Sustainable Urban TransportationBook from Wiley on AI for smart and sustainable urban transportation.

15 entries

Related directories on this site.

36Edge / Fog ComputingHub of edge computing directories. 37Physical AIRobot foundation models, simulation, edge compute and robots. 38Edge-Based Data Analytics SolutionsEdge analytics platforms and runtimes. 39Edge Kubernetes PlatformsLightweight Kubernetes and fleet management for edge sites. 40Edge Management and Orchestration (EMO) SoftwareSoftware for managing distributed edge infrastructure. 41AI Chip ProvidersProviders of AI chips and accelerators. 42Industry 4.0Industry 4.0 resources. 435G RAN Products5G radio access network products. 44Computing ParadigmsDirectory of computing paradigms. 45Cloud-Native Edge ComputingRunning cloud-native workloads at the edge. 46Edge Device Cloud PlatformsCloud platforms for managing edge devices. 47Edge Streaming Data AnalyticsStream processing and analytics at the edge. 48Serverless Edge PlatformsFunctions and compute at CDN and IoT edge locations. 49CPU, GPU, TPU, NPU, LPU, DPU & QPU ExplainedWhat each processor type is built for, including NPUs for on-device inference, and how to choose hardware. 50Modular, Scalable, AI-Ready Infrastructure for CPGWhy consumer goods manufacturers need it, with edge and cloud architecture.

Frequently Asked Questions

What is edge AI?

Edge AI runs machine learning models on devices or local servers near where data is produced, such as cameras, sensors and gateways, instead of sending the data to a central cloud for inference.

Why run AI at the edge?

It cuts latency, reduces bandwidth, keeps working without connectivity and can keep sensitive data on the device, which helps with privacy.

How do I get a model to run on a small device?

Models are usually compressed (quantised, pruned or distilled) and exported to a runtime such as LiteRT, ONNX Runtime, ExecuTorch or OpenVINO, or compiled for the target chip with tools such as Apache TVM or TensorRT.

Can language models run at the edge?

Yes, smaller or quantised models run locally with tools such as llama.cpp, Ollama and MLC LLM. Capability depends on the memory and accelerator available on the device.

Which entries changed or were left out?

Intel's OpenVINO moved to versioned documentation, and TensorFlow Lite is now LiteRT. Hailo and the Raspberry Pi AI Kit blocked automated checks, and the NXP eIQ and STM32Cube.AI links did not load, so they are not listed.