xnnpack

Prebuilt binary for TensorFlowLite's standalone installer. For RaspberryPi. A very lightweight installer. I provide a FlexDelegate, MediaPipe Custom OP and XNNPACK enabled binary.

🚀Explore how PyTorch models can be exported and lowered to the ExecuTorch Runtime for more effective, lightweight, minimalist inference at the Edge. Covers deploying to CPUs, particularly via the XNNPACK backend, for transformer-based and CNN models, before covering deployment to the Ethos-U backend, the TOSA IR, FVPs, and Google Model Explorer.🚀

100% offline ARM optimized Epcitetus persona chatbot

Embedded-first C++26/C23 neural inference: float32+int8 on MCU/MPU/CPU. CMSIS-NN · ESP-NN · NMSIS-NN · XNNPACK. Peer-benched vs TFLM, microTVM, TF Lite, ONNX Runtime (Arm · Espressif · RISC-V · Apple).

Android-first React Native ONNX Runtime bindings with Nitro Modules, backend selection, tensor inference, benchmarking, and experimental QNN diffusion pipeline helpers.

Matsuba vs benizuwai crab classifier from photos: pure-Go training, TFLite flatbuffer, XNNPACK inference

Universal AI Quantization Engine (UAQE) — an autonomous, capability-driven platform for AI model inspection, quantization, optimization, benchmarking, validation, and deployment on resource-constrained hardware.

Read くずし字 with a CNN: pure-Go training, hand-built CONV_2D/MAX_POOL_2D flatbuffer, XNNPACK inference

A drowsiness detection system using facial landmarks and hand detection to monitor user alertness. 😴📹🔔👁️

Iris species classifier: pure-Go training, TFLite flatbuffer, XNNPACK inference

Guess which 文豪 wrote a sentence: Aozora Bunko scraper, pure-Go training, TFLite flatbuffer, XNNPACK inference

Serve ExecuTorch .pte language models from an Android phone: an OpenAI and Anthropic compatible HTTP server that runs in the background, with KV cache reuse, run history and metrics. Kotlin Multiplatform core, Compose console.

Read handwritten digits with a CNN: pure-Go training, hand-built CONV_2D/MAX_POOL_2D flatbuffer, XNNPACK inference

Hardware-bound, deterministic AI runtime for edge devices (SM8750-AB) featuring thermal-aware manifold inference and Veritas Gate governance.

Offline acoustic machine-health monitoring on Arm. FP32→INT8: 4.07× faster in the deployed path, 3.68× smaller, accuracy retained. Every number regenerates from committed benchmarks.

Sazae-san janken predictor: scraper, pure-Go training, TFLite flatbuffer, XNNPACK inference

Benchmarks for compiled ExecuTorch models on real phones in AWS Device Farm: speed, memory, heat and answer quality, single and multi turn, every number traceable to a raw device log

On-device visual memory for the things you leave behind — Arm64 Android, EdgeTAM and ExecuTorch.

Photo Insight Android example application for on-device image classification and image-text matching with LiteRT, ExecuTorch, and XNNPACK

Android starter application for on-device image segmentation and depth estimation with ExecuTorch and XNNPACK