Desert Ant Labs has set its sights on placing a “little brain” inside every app.
The founders of a fresh European AI laboratory are placing their wager on a particular vision for the future of artificial intelligence. They believe it will be compact, quick, and run solely on a personal device carried by the individual.
Today marks the launch of Desert Ant Labs, which has released 18 on-device models built for audio, vision, and text work. The firm contends that most AI functions do not require a large cloud-based frontier model. Instead, they argue, a tailored model that operates locally, delivers responses within milliseconds, and charges nothing per call is sufficient. The initial 18 models are already available, with 12 stable and six in beta, all made accessible through a single SDK for Swift, Kotlin, and JavaScript.
The lab’s argument rests on a single idea: end API call payments when a small model running on the device delivers a superior result more quickly. The models carry no charge up to 100k monthly active devices, require no tokens, and need no logins at all.
The first 18 models and what they do
Desert Ant Labs ships one model per task, each designed to be the quickest method for completing that job on a device. The launch post names four flagship models.
- Voz: transcribes 10 minutes of audio in two seconds on an iPhone, which the company says is 4.7x faster than Whisper, with a start and end time on every word.
- Clear: 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 the data never reaches your servers.
- Tongue: identifies 84 languages from just three words, using a 2MB model.
A complete set of specifications and performance results for all fourteen models is posted on desertant.com/models and Hugging Face.
According to the company, the models are compact enough to operate on a device from five years ago and quick enough to work with every frame or keystroke. It is being claimed that they surpass the API call you are already paying for.
Why the founders built their own models
For five years, Desert Ant Labs worked on a video app called Detail, taking an on-device first approach. They did not begin as a model company. When they added features like Auto Edit for short clips or audio enhancement for podcasts, however, they found themselves relying on cloud APIs again. As Detail’s audience grew, so did their infrastructure costs.
According to the founders, they devoted years to tracking down models that work on-device, scanning Hugging Face for tools capable of spotting filler words or cleaning up a recording. Everything needed was present — the hardware, Core ML, the research — yet the distance between that groundwork and delivering a working feature proved too large to bridge.
The company trained its own models, treating the task as a product design challenge, since product is what they know. They built models and local inference that they say best cloud services on speed, quality, and cost, at a fraction of the size.
The company says it swapped Dolby’s audio processing for Clear, sped up its on-device transcriptions by 5 times using Voz, and traded Claude Sonnet for Clips, a 284MB model that breaks a 10-minute video into a dozen clips within five seconds.
The stated comparison puts Clips at a claimed speed advantage over Sonnet, alongside an energy efficiency edge, while keeping image quality unchanged. Specifically, they say Clips operates 10 times faster and consumes 470 times less power than Sonnet.
The detail 6, which arrives alongside iOS 27, swaps out every single cloud API for the company’s own models, and runs them entirely on the device itself.
The case against large language models
Desert Ant Labs is pushing back against the industry’s fixation on generalist frontier models. The company contends that developers have spent years treating LLMs as though they were merely another API, and in all the excitement over “frontier brains,” the industry nearly forgot that they are not the only choice available.
Every developer has a wish list of on-device models they would build if cost were not an issue, or features they are spending tokens on that they would gladly trade for a local model instead. The company describes a request that repeats itself a hundred thousand times a day: cleaning up a recording, labeling a photo, extracting a date from a sentence, or spotting a name before the text ever reaches your servers. None of these tasks requires a frontier model.
Desert Ant Labs points to NVIDIA’s own researchers, who broke down three agent systems and figured that roughly 40 to 70% of those systems’ calls to a large model could be handled by a smaller, specialized one instead.
The figure stands out for a coding group long accustomed to being charged by the token for each AI feature.
The economics of on-device compute
The financial argument is central to the launch. The industry will spend about $450 billion on data centers this year, according to Desert Ant Labs. Meanwhile, the world ships more than a billion phones, tablets, and laptops with increasingly capable chips, perfectly suited to these kinds of tasks.
The company’s position is that there is more compute available in people’s hands than in every AI data center on earth.
“We have an unfair advantage with free inference,” the founders write. No per-call cost means a feature can run on every message instead of only the ones you can afford to check. No round-trip means no latency. And your customer’s data never leaves the device.
According to the company’s argument, when inference costs nothing at all, the entire approach to building products shifts.
Building from the “little brain” up
Desert Ant Labs is designing its future on-device intelligence around a two-part architecture. The initial hundred models make up the cerebellum, “little brain.” The smaller brain takes care of constant tasks like balance and timing, which require little thought, leaving the rest of the brain free to handle more complex thinking.
The company’s initial effort centers on fast, specialized models designed to handle tasks that run continuously throughout the day, operating directly on the device without charge.
The cortex is the part that picks which model answers. It starts with a small local model, moves to a larger one when needed, and goes to the cloud only if the job demands it leave the device. As open research moves forward and device chips become more capable, the local models get bigger, and Desert Ant Labs says it will train larger ones itself. Frontier intelligence, built from the smallest models upward.
The firm criticizes cloud labs for shipping “neutral” models, claiming per-token pricing demands a neutral model. Each Desert Ant model arrives with a default selection set by the company, along with the controls developers can use to alter that default.
The models get tuned alongside the runtime itself. When running on an iPhone, both Clear and Voz use the Neural Engine. Over in the browser, Clear’s very same weights pass through WebAssembly instead.
Why Europe matters for data privacy
Desert Ant Labs is building in Europe, where “on-device” is the sovereign default. The company frames this as a privacy advantage: the data never leaves your customer’s hands, the feature never depends on someone else’s cloud, and what has never been uploaded can never be compelled.
The company’s Redact model takes this approach head-on. It hides names, addresses, and card numbers on the spot, doing so across 27 languages, which means the sensitive text never makes it to your servers at all.
The developer experience push
Desert Ant Labs argues that local models will gain traction only when the tools for building them improve dramatically. Developers need commercial-grade models that outperform the existing alternatives in both speed and quality, and they need to be able to integrate these models into an app with just a few lines of code. Ease of discovery is also essential, so developers can find the right model for their project quickly.
To address the discovery problem, the company posts complete specifications and performance numbers both on its own website and on Hugging Face. The software development kit supports three languages: Swift, Kotlin, and JavaScript.
The design of the pricing model aims to eliminate friction. Up to 100k monthly active devices are free, along with no tokens and no logins required. That means a developer can release a feature without having to build a billing integration or a signup flow.
What this means for the AI market
Desert Ant Labs is joining a field already packed with competitors, yet it has chosen its own path. While most AI labs push toward ever larger models, this one is moving in the opposite direction, racing instead to build models that are smaller and more focused.
The firm contends the processing power is already covered — it resides in the hands of a billion users. The challenge lies in crafting the software that puts that processing power to use.
It is not yet clear whether Desert Ant Labs can back up its promises regarding speed and accuracy. The firm contends that its models outperform cloud services on the task in question, while occupying a fraction of their space. Published benchmarks exist, though outside confirmation will require patience.
The wager at hand is straightforward: the path forward for AI does not lie with a larger brain in the cloud. Instead, it rests upon a small brain built into every product, operating without charge, on hardware that people already possess.
KEY FACTS BOX
- 18 models launched today: 12 stable, six in beta
- Free up to 100k monthly active devices
- Voz: transcribes 10 minutes of audio in two seconds on iPhone, claimed 4.7x faster than Whisper
- Clear: 9MB model, turns five-minute laptop recording into studio quality audio in one second
- Redact: masks data in real time in 27 languages
-
Tongue: 2MB model, identifies 84 languages from three words
-
Clips: 284MB model, turns 10-minute video into a dozen clips in five seconds, claimed 10x faster and 470x less energy than Claude Sonnet
- NVIDIA researchers estimated 40 to 70% of large model calls could go to small specialized models
- Industry projected to spend about $450 billion on data centers this year
- Detail 6 launches with iOS 27, replacing all cloud APIs with on-device models
- SDK supports Swift, Kotlin, and JavaScript
Source: desertant.com
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.