Desert Ant Core

// Overview

Desert Ant Core is the SDK layer for Desert Ant Labs, a European lab that publishes small, single-purpose models meant to run on the user's own device. Rather than one general model behind an API, the catalogue is a set of narrow tools: Voz transcribes speech, Clear denoises and dereverbs audio, Redact finds and masks personal data, Ear and Tongue identify a language from audio or from a few words of text, Gist tags content by topic, Clips pulls highlights out of long media, Uhm removes filler words, Title suggests titles and descriptions, Shapes cleans up rough sketches, and Emo suggests emoji. Six further models — Eye, Face, Moderator, Schemer, Toxic and Who — are in closed beta.

The packaging is the point. Each model is compiled for the inference runtime the platform already ships, so an app adds a dependency rather than an engine: Core ML on Apple platforms, LiteRT on Android, WebAssembly with LiteRT.js in the browser, and a prebuilt native core on Node. Products are per-model, so a project that only needs PII masking pulls Redact alone instead of a combined bundle. Everything runs locally, which means no request leaves the device, the feature keeps working offline, and there is no per-call token cost.

The repository is Swift-first with Kotlin and JavaScript SDKs alongside it, and the model weights are published separately on Hugging Face under the `desert-ant-labs` organisation. Licensing is source-available rather than OSI open source: the Desert Ant Labs Source-Available License is free for most apps, with a commercial licence required at scale, and GitHub reports the repository licence as non-standard. Desert Ant Labs states the models are free up to 100,000 monthly active devices per platform, with no cap on how often each person runs one.

// What it does

• Per-model packages rather than one large SDK — install only the models an app actually uses
• Core ML on iOS 18+ and macOS 15+, LiteRT on Android API 24+ (arm64-v8a and x86_64), WebAssembly with LiteRT.js in the browser, prebuilt native core on Node
• Native Swift, Kotlin and JavaScript APIs over the same model set
• Fully offline inference — audio, text and images never leave the device
• Audio models: Voz (transcription with word-level timing), Clear (denoise and dereverb), Ear (language ID from audio), Uhm (filler-word removal)
• Text models: Redact (PII detection and masking in 27 languages), Tongue (language ID from three words, 84 languages), Gist (topic tagging), Title, Emo
• Media models: Clips (highlight extraction), Shapes (single-stroke shape recognition), Align (word timestamps for an existing transcript)
• Weights published on Hugging Face under the desert-ant-labs organisation

// Getting started

Each platform installs the models it needs as individual packages. The examples below come from the repository README.

.package(url: "https://github.com/Desert-Ant-Labs/desert-ant-core.git", from: "3.1.0")

dependencies {
    implementation("ai.desertant:emo:3.1.0")
    implementation("ai.desertant:redact:3.1.0")
}

npm i @desert-ant-labs/emo @litertjs/core

let suggestions = try await Emo().suggestions(for: "Pay my bills")
let clean = try await Redact().redaction(of: "Email Anna at anna@example.hu.")

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Reach for it when a mobile or web app needs transcription, language detection or PII masking without a per-call API bill
• Reach for it when user audio, text or photos must not leave the device for privacy or compliance reasons
• Reach for it when a feature has to keep working with no network connection
• Reach for it when the job is narrow enough that a small task model beats calling a general frontier model

// How Desert Ant Core compares

Desert Ant Core alongside other open-source local runtimes tools AI/TLDR tracks, ranked by GitHub stars.