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

Desert Ant Labs launched today, describing itself as "a European frontier AI lab building opinionated on-device intelligence." The company says the best path to efficient intelligence starts on-device, and it is building small, specialized models for audio, vision and text that answer in milliseconds and cost nothing to run, small enough to work on a five-year-old phone and fast enough to use on every frame or keystroke. The first 18 models are live: 12 stable and six in beta, all reachable through one SDK for Swift, Kotlin and JavaScript. Every model is free up to 100,000 monthly active devices, with no tokens and no logins required.

Four models are named in detail. Voz transcribes 10 minutes of audio in two seconds on an iPhone, which Desert Ant Labs says is 4.7 times faster than Whisper, and it marks 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 across 27 languages so they never reach a developer's servers. Tongue identifies 84 languages from as little as three words, using a 2MB model. Desert Ant Labs says specs and benchmarks for the other 14 launch models are published on its own site and on Hugging Face.

The company frames the launch as five years in the making. It has been building a video app, Detail, with an on-device-first approach, but features like Auto Edit and audio enhancement kept forcing a fallback to cloud APIs, and Detail's infrastructure bills grew along with its popularity. Unable to find enough ready-made on-device models, Desert Ant Labs says it trained its own: it replaced Dolby with its own Clear model for audio enhancement, and it says Voz made its on-device transcriptions five times faster. It also says it replaced Claude Sonnet with Clips, a 284MB model that turns a 10-minute video into about a dozen clips in five seconds, which it describes as 10 times faster and using 470 times less energy than Sonnet at the same quality. The next version of the app, Detail 6, is set to launch alongside iOS 27 and will run entirely on Desert Ant Labs' own on-device models instead of any cloud API.

Desert Ant Labs argues that after years of treating large language models as the default API for everything, many everyday, high-frequency tasks such as cleaning up a recording, tagging a photo, pulling a date out of a sentence, or catching a name before text reaches a server do not need a frontier model at all. It cites NVIDIA's own 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 model instead. The company also frames on-device compute as already paid for: it says the industry will spend about $450 billion on data centers this year, while the world ships more than a billion phones, tablets and laptops with increasingly capable chips, which it argues puts more compute in people's hands than in every AI data center on earth.

Desert Ant Labs lays out a roadmap in two stages. It calls its first hundred models the "cerebellum," the tier it says it is building now to handle always-on, low-level work for free, and says a later "cortex" layer will decide which model answers a given request: a small local model first, a bigger one when the job needs it, and the cloud only when work has to leave the device. As device chips grow more capable, the company says it will train larger local models itself, working toward what it calls "Frontier intelligence, built from the small end up." It also contrasts its approach with cloud labs, which it says ship neutral models because per-token pricing requires neutrality; every Desert Ant model instead ships with a default the company chooses, plus the levers to change it. The SDK is available now on GitHub for Swift, Kotlin and JavaScript, with documentation the company says is written for both developers and AI agents; models can also be tried through a command-line tool on a Mac or directly in a browser on Hugging Face.

Key facts

Desert Ant Labs launched today with 18 on-device models (12 stable, six in beta) for audio, vision and text, accessible through one SDK for Swift, Kotlin and JavaScript, free up to 100,000 monthly active devices.
Voz transcribes 10 minutes of audio in two seconds on an iPhone, which the company says is 4.7 times faster than Whisper; Clear, a 9MB model, turns a five-minute recording into studio-quality audio in one second; Redact masks personal data in real time across 27 languages; Tongue identifies 84 languages from three words with a 2MB model.
The company says its 284MB Clips model turns a 10-minute video into about a dozen clips in five seconds, replacing Claude Sonnet in its own video app Detail at 10 times the speed and 470 times less energy, at the same quality.
Citing NVIDIA's own researchers, who examined three agent systems, Desert Ant Labs says 40 to 70% of calls to a large model in those systems could go to a small, specialized model instead.
Detail 6, launching alongside iOS 27, is set to replace all of Desert Ant Labs' cloud APIs with its own on-device models.

Why it matters

Desert Ant Labs is making a specific bet: that years of treating large language models as the default API for every feature have obscured a cheaper option for small, repeated tasks. Its argument is that transcription, audio cleanup, language detection and similar jobs run the same way thousands of times a day and need speed and low cost, not a frontier model's general reasoning. It backs this with a citation to NVIDIA's own researchers, who examined three agent systems and estimated that 40 to 70% of their large-model calls could go to a small, specialized model instead. The company also reframes the cost side of the AI buildout: against an industry it says will spend about $450 billion on data centers this year, more than a billion phones, tablets and laptops already ship with capable chips, compute it argues is effectively already paid for and sitting idle in people's pockets.

Who it affects

The immediate audience is app developers who currently pay per call for tasks like transcription, audio cleanup, personal-data redaction or language detection, and who could move some of that work to a free, on-device model instead. Desert Ant Labs points to its own video app, Detail, as the proof case: it says it replaced Dolby with its Clear model for audio enhancement and replaced Claude Sonnet with its own Clips model for turning long recordings into clips, and the next version, Detail 6, is set to run entirely on Desert Ant Labs' own models when it launches alongside iOS 27. The company also frames this as an appeal to European developers specifically: it says it is building in Europe, where "on-device" is the sovereign default, so a feature never depends on someone else's cloud and data that is never uploaded can never be compelled.

How to use it

The SDK is native for Swift, Kotlin and JavaScript and available now on GitHub, with documentation the company says is written for both developers and AI agents. Every model is free up to 100,000 monthly active devices, with no per-token charges and no login required; the source does not state what usage above that threshold costs. Developers can try the models before integrating them, either through a command-line tool on a Mac or directly in a browser on Hugging Face. Four models are detailed with names and numbers in the launch post: Voz for transcription, Clear for audio enhancement, Redact for masking personal data, and Tongue for language identification; specs and benchmarks for the other 14 are published on Desert Ant Labs' own site and on Hugging Face.

How solid is it

Every figure in the launch comes from Desert Ant Labs itself: the comparisons against Whisper, Dolby and Claude Sonnet, and the timing and size numbers for each model, are the company's own reported benchmarks, with no outside reviewer or independent test cited. No founder, executive or spokesperson is named anywhere in the post, even though the "How we got here" section is written in the first person. The one outside data point, NVIDIA researchers' estimate that 40 to 70% of large-model calls could go to a smaller model, is reported secondhand, with no paper, citation or link given for it. The post also states two different speed figures for Voz's transcription: 4.7 times faster than Whisper by name in the model list, and a separate five times faster with no baseline named in the origin story, without saying whether the two describe the same comparison.

Risks and caveats

The launch numbers are all self-reported and unaudited, and several details a buyer would want are simply not in the post: no price for usage above the 100,000-monthly-active-device free tier, no funding, investor or headcount information, no city or country within Europe for the company itself, and no specific version named for either Whisper or Claude Sonnet in the comparisons. The product line is also new: six of the 18 launch models are still in beta, and the roadmap beyond the first hundred models depends on device silicon getting more capable, a trend the company assumes without committing to a timeline. The claim that people's devices hold more compute than every AI data center on earth is asserted without any supporting figures.

“We have an unfair advantage with free inference. No per-call cost, so a feature runs on every message instead of the ones you can afford to check. No round-trip, and your customer's data never leaves the device.”

— Desert Ant Labs