How much does Desert Ant Labs cost?

Desert Ant Labs is free up to 100,000 monthly active devices per platform, with no tokens and no logins, and there is no limit on how often each person runs a model. Past that scale the code falls under the Desert Ant Labs Source-Available License, which requires a commercial licence arranged with the lab.

Which platforms and versions do Desert Ant models support?

Desert Ant models run on Apple platforms from iOS 18 and macOS 15 upward through Core ML, on Android from API 24 upward through LiteRT for arm64-v8a and x86_64, in the browser through WebAssembly with LiteRT.js, and in Node through a prebuilt native core. Swift, Kotlin and JavaScript SDKs cover those targets.

Is Desert Ant Labs open source?

Desert Ant Labs publishes its SDK code on GitHub and its weights on Hugging Face, but ships under the Desert Ant Labs Source-Available License rather than an OSI-approved open-source licence, which GitHub reports as a non-standard licence. Most apps can use it for free; larger deployments need a commercial licence from the lab.

Which Desert Ant models are still in beta?

Six of the 18 Desert Ant models are in beta: Eye ranks images by aesthetics and content, Face finds people in photos on-device, Moderator flags nudity, Schemer extracts typed JSON from text, Toxic catches hate speech, and Who labels speakers in audio and video. The other 12, including Voz, Redact and Clear, are listed as available.

How does Voz compare with Whisper?

Voz transcribes ten minutes of audio in two seconds on an iPhone, which Desert Ant Labs measures as 4.7x faster than Whisper, and it returns word-level timing. Voz also runs entirely on the device, so recordings never reach a server — a different trade-off from cloud Whisper APIs that bill per minute of audio.

Desert Ant Labs · 2026-09-08 · major

Desert Ant Labs — 18 small AI models that run offline inside your app

Desert Ant Labs launched 18 on-device models with Swift, Kotlin and JavaScript SDKs. Each does one job — transcription, PII masking, language detection, video clipping — and runs fully offline. Free up to 100,000 monthly active devices.

Desert Ant Labs launch announcement for on-device AI models

Eighteen small models that run fully offline inside iOS, Android and web apps, with no token bill.

Quick facts

Maker	Desert Ant Labs
Models	18 (12 stable, 6 beta)
SDKs	Swift, Kotlin, JavaScript
Runtimes	Core ML, LiteRT, WebAssembly
Free tier	100k monthly active devices
License	Source-available
Announced	8 September 2026

Benchmarks

PII detection accuracy

Redact (12MB)		88.8%
GLiNER-PII (2.3GB)		91.1%

source ↗

What is it?

Desert Ant Labs launched 18 small on-device models on 8 September 2026 — 12 stable and six in beta — each doing one narrow job. Voz transcribes speech, Redact masks personal data, Tongue detects a language from three words, Clips cuts a long video into shorts, Uhm strips filler words. The lab describes itself as a European lab building opinionated on-device models and ships them behind native Swift, Kotlin and JavaScript SDKs.

How does it work?

Each model is compiled for the runtime a platform already ships, so there is no separate inference engine to bundle. Apps pull one package per model instead of a single large SDK — a project that only needs PII masking installs the 12MB Redact model and nothing else. Inference happens on the device, so audio and text never leave it and the feature keeps working with no network.

Why does it matter?

The everyday jobs inside an app — detect the language, strip the personal data, cut the clip, suggest a title — are the ones teams quietly pay frontier-model prices for, per call, forever. Desert Ant Labs turns those into a local dependency that costs nothing at ordinary scale and removes the privacy review that comes with sending user audio to a third party. The trade is accuracy: the small models sit a few points behind the large cloud ones.

Who is it for?

mobile and web app developers

Frequently asked questions

How much does Desert Ant Labs cost?: Desert Ant Labs is free up to 100,000 monthly active devices per platform, with no tokens and no logins, and there is no limit on how often each person runs a model. Past that scale the code falls under the Desert Ant Labs Source-Available License, which requires a commercial licence arranged with the lab.
Which platforms and versions do Desert Ant models support?: Desert Ant models run on Apple platforms from iOS 18 and macOS 15 upward through Core ML, on Android from API 24 upward through LiteRT for arm64-v8a and x86_64, in the browser through WebAssembly with LiteRT.js, and in Node through a prebuilt native core. Swift, Kotlin and JavaScript SDKs cover those targets.
Is Desert Ant Labs open source?: Desert Ant Labs publishes its SDK code on GitHub and its weights on Hugging Face, but ships under the Desert Ant Labs Source-Available License rather than an OSI-approved open-source licence, which GitHub reports as a non-standard licence. Most apps can use it for free; larger deployments need a commercial licence from the lab.
Which Desert Ant models are still in beta?: Six of the 18 Desert Ant models are in beta: Eye ranks images by aesthetics and content, Face finds people in photos on-device, Moderator flags nudity, Schemer extracts typed JSON from text, Toxic catches hate speech, and Who labels speakers in audio and video. The other 12, including Voz, Redact and Clear, are listed as available.
How does Voz compare with Whisper?: Voz transcribes ten minutes of audio in two seconds on an iPhone, which Desert Ant Labs measures as 4.7x faster than Whisper, and it returns word-level timing. Voz also runs entirely on the device, so recordings never reach a server — a different trade-off from cloud Whisper APIs that bill per minute of audio.

Try it

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