Success Stories#
See where ExecuTorch is shipping today, how organizations integrate it, and what current reference implementations demonstrate. Labels on this page are intentional: production claims link to primary sources, while prototypes and experimental backends are identified as such.
Production Deployments#
Status: Production
Platforms: Android and iOS
Scale: Features serving billions of people
ExecuTorch runs on-device models behind Instagram Cutouts, WhatsApp bandwidth estimation, Messenger language identification and encrypted experiences, and Facebook Stories music recommendations.
Status: Production
Devices: Meta Quest 3 and 3S; Ray-Ban Meta, Oakley Meta Vanguard, and Meta Ray-Ban Display glasses
ExecuTorch powers capabilities including hand and controller tracking, persistent room memory, live translation, visual captions, and on-device OCR.
Status: Production
Platforms: macOS and Windows
Model: NVIDIA Parakeet TDT
LM Studio ships local voice transcription powered by ExecuTorch. The same model and application layer target Apple GPUs on macOS and NVIDIA GPUs on Windows.
Customer and Product Case Studies#
Status: Published customer case study
Hardware tested: AMD Ryzen AI 9 HX 370 and Samsung Galaxy S24 CPUs
Reported result: Up to 2× higher CPU throughput than selected similarly sized models, with reduced memory use
Liquid AI adopted ExecuTorch for LFM2 models from 350M to 4B parameters. Its case study highlights support for hybrid attention and recurrent architectures, portable model packaging, and integration with the LEAP platform.
Status: Shipping and open source
Platforms: iOS and Android
Workloads: Chat, document retrieval, images, and speech input
Software Mansion built Private Mind with React Native ExecuTorch. After models are downloaded, conversations, retrieval, embeddings, and inference stay on the device.
Ecosystem Integrations#
Transformers includes an experimental generic ExecuTorch exporter, Optimum ExecuTorch adds tested task-level recipes and model wrappers, and torchao supplies PyTorch-native quantization recipes and kernels used throughout ExecuTorch.
Unsloth documents a workflow for fine-tuning LLMs with torchao quantization-aware training, then exporting them to ExecuTorch for local Android and iOS inference.
A React Native library with task-level APIs, a pre-exported model catalog, and XNNPACK, Core ML, MLX, and Vulkan acceleration.
Ultralytics provides a first-class ExecuTorch export path for YOLO26 across its seven vision tasks, targeting mobile and edge applications.
Digica’s AI SDK automates model conversion and generates C++, Java, and Objective-C++ interfaces for ExecuTorch. Its published mobile demo combines face detection, alignment, recognition, and emotion classification.
The open-source DeliteAI SDK offers ExecuTorch as one of its model runtimes for Python-orchestrated agent workflows in Android and iOS applications.
NVIDIA FLARE uses ExecuTorch for local training in federated-learning workflows, with Android and iOS SDK examples and real-device and simulated setups.
The Arm ML Embedded Evaluation Kit supports Cortex-M and Ethos-U targets. Alif has demonstrated generative AI and real-time speech-to-text on its Ensemble E8, and the ExecuTorch Arduino library is hardware-verified on Arduino UNO Q.
The OpenVINO delegate supports Intel CPU, GPU, and NPU deployment, with end-to-end examples for YOLO26, Llama, and Stable Diffusion.
Technical Showcases#
These are reference implementations, not claims of production deployment. Performance figures apply only to the linked model, device, and measurement setup.
Browse representative model examples →
Maturity: Reference implementation
Platforms: NVIDIA CUDA and Apple silicon through MLX
Run text and image input, 128K-token context, OpenAI-compatible serving, GGUF K-quant weights, and DFlash speculative decoding. The published M5 Pro test improved decode from 21.6 to 33.0 tokens/s (52.8%) without quality regression.
Maturity: Reference implementation
Device measured: Samsung Galaxy S25
Gemma 4 E2B and E4B examples combine audio transcription, translation, image understanding, and text generation. The E2B 4-bit audio configuration reports a 0.71 real-time factor, 6 tokens/s generation, and 2,251 MB peak memory on the documented 23-second sample.
Maturity: Reference implementations; one production adopter
Build transcription, streaming ASR, diarization, and voice activity detection with Parakeet TDT, Voxtral Realtime, Whisper, Sortformer, and Silero VAD across CPU, GPU, and NPU backends.
Maturity: Reference implementations
Voxtral TTS 2603 supports CPU and CUDA; its documented RTX 5080 configuration runs at 0.31× real-time factor, more than 3× real time. Its model weights and voice embeddings are CC BY-NC 4.0. Supertonic 3 adds a dynamic FP16 text-to-speech path for the experimental MLX delegate.
Maturity: Early model support
The dense 0.8B, 2B, and 4B path currently targets FP32 static-shape XNNPACK. The 35B-A3B mixture-of-experts reference adds INT4 export, CUDA and MLX runners, and OpenAI-compatible serving. Review each example’s limitations before use.
Maturity: Experimental
The WebGPU backend demonstrates language, vision, retrieval, audio, and on-device training workflows. A July 2026 M4 Pro measurement reports 188.3 decode tokens/s for a 4-bit Llama 3.2 1B artifact at 128-token context.
More reference workflows:
LoRA weight sharing in C++: share a foundation-weight file across model programs and reuse those weights through XNNPACK’s weight cache.
Stable Audio Open Small on Arm CPUs: generate audio from text prompts with ExecuTorch, XNNPACK, and Arm KleidiAI. Arm and Meta video session →
Community Showcase#
At the June 2026 ExecuTorch Hackathon, more than 100 participants across over 20 teams built on-device prototypes on Snapdragon-powered Samsung Galaxy S25 Ultra phones. The winning prototypes explored real-time visual safety, haptic navigation for blind and low-vision users, and private gait analysis. These are community prototypes, not production deployments.
NoemaAI
Status: Community app; not yet vetted by the ExecuTorch team.
NoemaAI’s documentation and App Store release notes list ExecuTorch as one of the runtimes available in its Apple-platform AI app.
Runtime documentation → • App Store →
Want your project considered for this page? Submit a success story with its deployment status, model, device and backend, reproducible measurements and baseline, a public primary source, and permission to use any supplied logo or image.
