Desert Ant Labs launches 18 on-device AI models with one SDK

Desert Ant Labs, a European frontier AI lab, launched on September 8, 2026 with eighteen specialized on-device models for audio, vision and text, a single SDK for Swift, Kotlin and JavaScript, and a pricing model that makes local inference free.

One model per task

Twelve models are stable and six are in beta. Each answers in milliseconds and costs nothing to run, which makes it usable on every frame or keystroke and small enough for a five-year-old phone, the company says. Voz transcribes ten minutes of audio in two seconds on an iPhone — 4.7x faster than Whisper — with a start and end time on every word. Clear is a 9MB model that turns a five-minute laptop recording into studio-quality audio in one second. Redact masks names, addresses and card numbers in real time in 27 languages, so they never reach a server. Tongue identifies 84 languages from three words with a 2MB model, scoring 0.933 against 0.887 for a 293MB detector, while Redact catches 88.8% of personal data versus 91.1% for the 2.3GB GLiNER-PII.

Why on-device

The lab frames local inference as Europe's sovereign default: data never leaves the customer's hands, a feature never depends on someone else's cloud, and what was never uploaded cannot be compelled. The economics rest on compute that is already paid for: the industry will spend about $450 billion on data centres this year, while more than a billion phones, tablets and laptops ship with increasingly capable chips — more compute in people's hands than in every AI data centre on earth. Desert Ant also cites NVIDIA researchers who examined three agent systems and estimated that 40 to 70% of their calls to a large model could go to a small, specialized one instead. Every model is free up to 100,000 monthly active devices.

From a video app to its own models

The team spent five years building the video app Detail, but fell back to cloud APIs for features like Auto Edit clip creation and podcast audio enhancement as infrastructure bills grew. They trained the models themselves: Clear replaced Dolby and Voz made on-device transcription five times faster. Clips, a 284MB model that turns a ten-minute video into a dozen clips in five seconds, replaced Claude Sonnet at 10x the speed and 470x less energy with the same quality, the company says. Detail 6, due with iOS 27, will replace all of its cloud APIs with the same models running entirely on device.

Little brains first

Desert Ant describes the first hundred models as a cerebellum — the little brain handling always-on work so the rest is free to think — and a later cortex layer that decides which model answers: a small local model first, a bigger one when the job requires it, and the cloud only when work has to leave the device. Models ship with a chosen default plus levers to change it, and the runtime is optimized alongside: Clear and Voz run on the iPhone Neural Engine, while Clear's weights run through WebAssembly in the browser. The SDK is on GitHub, with docs, a Mac CLI and browser demos on Hugging Face.