Desert Ant Labs launches 18 on-device AI models for audio, vision and text

A European lab betting on local inference

Desert Ant Labs, a new European AI lab, has launched with 18 small, task-specific models covering audio, vision and text, all built to run entirely on the user's device. Twelve ship as stable releases and six in beta, reachable through a single SDK for Swift, Kotlin and JavaScript. According to the company's launch announcement, which reached the front page of Hacker News, the models are free for up to 100,000 monthly active devices, with no per-token pricing and no login required.

Inference with no marginal cost

The lab's core argument is that a large share of AI-assisted product features — tidying up a recording, tagging a photo, extracting a date from a sentence, masking a name before text reaches a server — does not need a frontier model. Per-call pricing and network latency, the company says, quietly decide which interactions actually get intelligence, and local models remove both constraints: responses arrive in milliseconds and each extra call costs nothing.

To support the case, the announcement cites NVIDIA researchers who analysed three agent systems and estimated that 40 to 70 percent of their calls to a large model could have been handled by a small, specialised one. It also contrasts the roughly $450 billion the industry will spend on data centers this year with the more than a billion phones, tablets and laptops shipped with silicon well suited to exactly these workloads.

The models and the company's benchmarks

Four models anchor the launch. Voz transcribes ten minutes of audio in about two seconds on an iPhone, which the company says is 4.7 times faster than Whisper, and attaches start and end times to every word. Over 30 minutes of audio on an M3 Ultra, Desert Ant Labs reports a realtime factor of 319 for Voz, against 78 for Apple SpeechAnalyzer and 50 for Whisper large-v3-turbo.

Clear is a 9MB model said to turn a five-minute laptop recording into studio-quality audio in one second, running at 302 times realtime on an iPhone 16 Pro and 345 times on a MacBook Pro with an M5 chip. Redact masks names, addresses and card numbers in real time across 27 languages from 12MB of weights, catching 88.8 percent of personal data in the lab's tests — close to the 91.1 percent it measured for GLiNER-PII, a 2.3GB model. Tongue identifies 84 languages from three spoken words using a 2MB model, scoring 0.933 accuracy against 0.887 for a 293MB detector.

A fifth model, Clips, was built to replace Claude Sonnet inside the founders' own product: at 284MB it reportedly turns a ten-minute video into a dozen clips in five seconds, which the company frames as ten times faster and 470 times more energy-efficient at equivalent quality. All of these figures are self-reported. Full specifications for the remaining models are published on the company's site and on Hugging Face, and the SDKs sit on GitHub alongside a Mac CLI for trying models locally.

Grown out of a video app's cloud bills

The lab emerged from Detail, a video app the team has built for five years with an on-device-first philosophy. Features such as automatic clip editing and podcast audio enhancement kept falling back to cloud APIs, and as Detail grew more popular, so did its infrastructure bills. Periodic searches of Hugging Face for deployable local models never surfaced anything the team could drop into a shipping product, so they concluded the missing piece was product design rather than research, and trained their own models. Detail 6, which launches alongside iOS 27, will replace every cloud API in the app with the lab's own models.

A "little brain" first, routing later

The roadmap is deliberately bottom-up. Desert Ant Labs describes its first hundred models as the cerebellum — the always-on layer handling skills users never think about — with a routing cortex to follow: a small local model answers first, a bigger one steps in when the job demands it, and the cloud is used only when work must leave the device. Because the lab does not sell tokens, it says it can ship opinionated defaults rather than neutral models, while exposing levers to change them. Runtime and weights are tuned together: on iPhone, Voz and Clear run on the Neural Engine, and in the browser Clear's weights run through WebAssembly.

Why it matters

Most AI roadmaps assume a generalist model behind an API. Desert Ant Labs is betting the opposite: that routine, high-frequency tasks belong on hardware users already own, where inference is free, latency is negligible and data never leaves the device — a framing with clear resonance in Europe, where data sovereignty rules. If the benchmarks hold up under independent testing, the consequence is product design freed from token budgets: a check can run on every message rather than the ones a team can afford. Until then, treat the numbers as launch-post marketing, and note that the company has not said what pricing applies beyond the 100,000-device free tier.