Desert-Ant-Labs/desert-ant-core

What it solves

Desert Ant Core provides a suite of small, focused AI models designed to run entirely on-device. This eliminates the need to send sensitive user data (text, audio, and images) to a cloud server, ensuring privacy and reducing latency by performing inference directly on phones, Macs, or in web browsers.

How it works

The project provides SDKs for Swift, Kotlin, and JavaScript/TypeScript. It leverages platform-specific hardware acceleration: Core ML for Apple devices, LiteRT (formerly TensorFlow Lite) for Android and other platforms, and WebAssembly with LiteRT.js for the web. Models are hosted on Hugging Face and can be downloaded and cached locally, with support for airgapped environments and custom model directories.

Who it’s for

Developers building privacy-first applications for iOS, Android, macOS, Windows, Linux, and the web who need specific AI capabilities like PII redaction, speech recognition, or content tagging without relying on cloud APIs.

Highlights

• Diverse Model Library: Includes models for PII redaction (Redact), emoji suggestions (Emo), speech recognition (Voz), language identification (Tongue/Ear), and speech enhancement (Clear).
• Cross-Platform Consistency: Models behave identically across all supported platforms, allowing for a single feature implementation to be shipped everywhere.
• On-Device Privacy: All processing happens locally; data never leaves the device.
• Flexible Deployment: Supports managed caches, self-hosted model files, and offline/airgapped installations.
• Broad Runtime Support: Compatible with Core ML, LiteRT, and WebAssembly.