Desert Ant Labs puts inference inside the loop

Desert Ant Labs says its 2MB language detector, Tongue, names a spoken language from three words with 0.933 accuracy, beating a 293MB detector that scores 0.887. The size-to-accuracy gap is the whole pitch. The European lab launched with 18 models, 12 stable and six in beta, each trained to do one task on a phone, tablet or laptop instead of in a data center.

The models cover audio, vision and text. Voz transcribes ten minutes of audio in two seconds on an iPhone, which the company puts at 4.7 times faster than Whisper, with a start and end time on every word. Clear, a 9MB model, turns a five-minute laptop recording into what Desert Ant calls studio-quality audio in one second, running at 302 times realtime on an iPhone 16 Pro and 345 times on an M5 MacBook Pro. Redact masks names, addresses and card numbers in 27 languages before text leaves the device, and the company says its 12MB version catches 88.8% of the personal data in a text against 91.1% for GLiNER-PII at 2.3GB. Everything runs through one SDK for Swift, Kotlin and JavaScript, and every model is free up to 100,000 monthly active devices, with no tokens and no logins.

One model per task is the design decision, optimized together with the runtime it ships on. On an iPhone, Clear and Voz run on the Neural Engine. In a browser, Clear’s same weights run through WebAssembly.

The lab came out of a video app, Detail, that the team built on-device-first for five years. Features like Auto Edit and audio enhancement kept forcing them back to cloud APIs, and the infrastructure bill grew with the user base. So they trained replacements: Clear for Dolby audio enhancement, Voz for transcription at five times the previous speed, and Clips, a 284MB model that turns a ten-minute video into a dozen clips in five seconds. The company says Clips replaced Claude Sonnet for that job at ten times the speed and 470 times less energy, with the same quality. Detail 6, launching with iOS 27, drops every cloud API in favor of these models running entirely on the device.

The argument underneath the launch is about where inference belongs. Desert Ant cites NVIDIA researchers who pulled apart three agent systems and estimated that 40 to 70% of their calls to a large model could have gone to a small specialized one instead. For work that runs the same way a hundred thousand times a day, cleaning a recording, tagging a photo, pulling a date out of a sentence, catching a name before it reaches your servers, a frontier model is overkill on cost, latency and energy.

The economics are the point. The company notes the industry will spend about $450 billion on data centers this year, while more than a billion phones, tablets and laptops ship with capable chips. When a call costs nothing per invocation and makes no round trip, you can run a check on every message rather than the subset you can afford to inspect.

For anyone with compliance exposure, the on-device model changes the threat surface. Data that never leaves the device cannot be compelled, logged by a third party, or leaked from someone else’s cloud. Redact running locally means PII is masked before it reaches your servers at all, which is a cleaner story for a data protection assessment than scrubbing records after ingestion.

If you want to try this, start by auditing per-call cloud spend for the high-frequency, low-variance work: transcription, audio cleanup, language detection, PII masking, photo tagging. Those are the tasks with a narrow definition and the volume to make the swap pay. Every benchmark in the launch is the vendor’s own, measured on hardware it chose, so test the size-to-accuracy claims against your own data before you remove a cloud dependency.

Desert Ant’s longer plan is a routing layer it calls the cortex: a small local model answers first, a bigger one when the job needs it, and the cloud only when the work has to leave the device. That layer is not shipped. What is live now is the first set of specialized models and the SDK on GitHub, with a CLI to run the models on a Mac and the weights on Hugging Face.