Eighteen models that run on a five-year-old phone

The benchmark the company puts forward is not size but response time. Paul Veugen, its founder and chief executive, wrote that "we're building small, specialized models for audio, vision, and text, and each model answers in milliseconds." The design condition is that a model has to run on a five-year-old phone and stay fast enough to be called on every video frame or every keystroke. Once the cost of a single call disappears, the company argues, a feature can run on every message rather than only on the ones you can afford to check.

The numbers are broken out task by task. Voz, the transcription model, moves 10 minutes of audio in two seconds on an iPhone and attaches a start and end time to every word. In measurements the company published, the realtime multiple over 30 continuous minutes of audio was 319 times for Voz, 78 times for Apple SpeechAnalyzer and 50 times for Whisper large-v3-turbo. On an iPhone 17 Pro it reached 298 times.

Clear, which cleans up audio, is a 9MB model. The company says it turns a five-minute laptop recording into studio quality in one second. It was measured at 302 times realtime on an iPhone 16 Pro and 345 times on an M5 MacBook Pro. The same weights run on the Neural Engine on an iPhone and through WebAssembly in the browser.

With Redact, which masks personal data, the company also printed the comparison it loses. At 12MB, Redact catches 88.8% of the personal data in a document, below the 91.1% of the 2.3GB GLiNER-PII. The 14.7MB Rampart managed 61.4% and a 3GB OpenAI filter 60.2%. Tongue, which identifies languages, distinguishes 84 of them from three words at 2MB and scored 0.933, against 0.887 for a 293MB detector.

The company swapped the models into its own app first. In Detail, the video app it has been building for five years, audio enhancement moved from Dolby to Clear and transcription to Voz, which is five times faster. Cutting a 10-minute video into a dozen clips is handled in five seconds by the 284MB Clips, and the company says it replaced Claude Sonnet in that slot, running 10 times faster and using 470 times less energy at the same quality. Detail 6, due alongside iOS 27, strips out cloud APIs entirely and runs on the device.

For an engineer, the heart of this launch is not model size but where calls get sent. The launch post METAL read in full cites an estimate from NVIDIA researchers who pulled apart three agent systems and concluded that 40 to 70% of their calls to a large model could go to a small, specialized one instead. Work that repeats a hundred thousand times a day, such as cleaning up a recording, tagging a photo, pulling a date out of a sentence or catching a name before the text reaches a server, does not need a frontier model. The company likens this layer to the cerebellum, which handles the always-on work, and says the cortex layer that decides which model answers comes next.

It also does the arithmetic on where compute already sits. While putting industry data-center spending this year at roughly $450 billion, the company wrote that "there's more compute available in people's hands than in every AI data center on earth." More than a billion phones, tablets and laptops ship every year with increasingly capable chips. The models can be used on up to 100,000 monthly active devices per SDK, with no cap on inference per user.

Where it is built is part of the design too, the company wrote. In Europe, running on the device is the default, and if data never leaves the customer's hands, a feature does not depend on someone else's cloud and what was never uploaded cannot be compelled. Attempts to put small models on devices have grown noticeably since the summer. METAL reported on Liquid AI open-sourcing Pipette, an evaluation suite for on-device models. What separates this launch from that trend is that it ships neither an evaluation tool nor a single model, but a set of task-specific models and commercial terms together.

In short, Desert Ant Labs has started with a bundle of 18 models, one per task, and a single SDK, and it proved the claim on itself first by stripping cloud calls out of its own app. The published numbers run well ahead on speed and size, and fall a little short of far larger models on accuracy. Two things are worth watching. What terms attach to apps that pass the 100,000 active-device mark, and whether the cortex layer the company has trailed, the one that decides which model gets the job, actually arrives.