QVAC vs ExecuTorch
QVAC
Open-source ecosystem for local-first peer-to-peer AI on every platform.
ExecuTorch
PyTorch's on-device runtime for mobile, embedded and edge hardware.
Key differences
ExecuTorch is PyTorch's runtime for edge hardware, running on Android, iOS, desktop and microcontrollers across more than twelve backends including CoreML, Qualcomm, MediaTek and Samsung Exynos. It supports on-device training, and its program-data separation lets LoRA adapters share one set of foundation weights. QVAC reaches on-device fine-tuning through Fabric, which contributes GPU-accelerated LoRA and masked-loss instruction tuning on mobile hardware.
The workflows and the languages differ. ExecuTorch is ahead-of-time: a PyTorch model is exported to a .pte file, then loaded from C++, Swift or Kotlin, with no JavaScript or TypeScript path. QVAC exposes JavaScript, TypeScript and Python, loads GGUF at runtime from a file, a URL or a peer, and reaches fine-tuning through a single call.
The two ship different amounts of the application layer. ExecuTorch provides a runtime, an export toolchain and example models to adapt. QVAC provides task-level APIs for transcription, translation, OCR, speech synthesis, embeddings, RAG and image generation, with a model registry behind them.
This page compares ExecuTorch v1.4.1, released 14 August 2026, against QVAC 0.18.2, meaning the SDK together with the Fabric inference engine at v10297.1.1. Every row was checked against the project's own documentation and release notes on 4 September 2026. Both projects move quickly, so check the current release before you make a decision on either one.
Feature matrix
Feature
QVAC
ExecuTorch
PLATFORMS
macOS
Windows
Linux
Android
iOS
AI TASKS
Text generation
Transcription
Translation
Image generation
OCR
Text-to-speech
RUNTIME SUPPORT
Node.js
Bare
Expo
HTTP server
CLI
Export only
P2P
Peer discovery
Inference delegation
Encrypted transport
MOBILE SUPPORT
On-device inference
LoRA fine-tuning on mobile
Mobile SDK
LICENSING
License
Apache 2.0
BSD-3-Clause
Open weights tooling
When to choose QVAC
Your application is JavaScript, TypeScript, React Native or Python.
You want models loaded at runtime, with no export step.
You need speech, OCR, translation or images supplied.
You want peer-to-peer model distribution and delegation.
When to choose ExecuTorch
Your models are PyTorch and your team ships native code.
You need a specific NPU backend such as CoreML or Qualcomm.
You are targeting microcontrollers or memory-constrained hardware.
You want ahead-of-time control over the whole graph.
Compare
Ready to build with QVAC?
One SDK, every platform, no rent. Grab it and ship your own local-first AI.
npm install @qvac/sdk