Today we're launching Desert Ant Labs: a European frontier AI lab...
Desert Ant Labs@desertantlabs
94 views
Sep 10, 2026
~2 min read
Advertisement
1
Today we're launching Desert Ant Labs: a European frontier AI lab building on-device intelligence.
18 models across audio, vision, and text. SDKs for Swift, Kotlin, and JavaScript.
No tokens. No logins. Nothing leaves the device.
desertant.com
18 models across audio, vision, and text. SDKs for Swift, Kotlin, and JavaScript.
No tokens. No logins. Nothing leaves the device.
desertant.com
2
There's more compute available in people's hands than in every AI data center on earth.
A billion devices a year ship with a chip built for exactly this work. Paid for at the point of sale, idle most of the day.
We put that compute to work.
A billion devices a year ship with a chip built for exactly this work. Paid for at the point of sale, idle most of the day.
We put that compute to work.
3
Voz: transcribe 10 minutes in 2 seconds on an iPhone, 4.7x faster than Whisper, with a start and end time on every word.
25 European languages. 467MB.
25 European languages. 467MB.
0:24
4
Clear: studio sound, no cloud bill.
A laptop mic in a room with traffic outside comes out close-miked. 5 minutes of audio in 1s, using a 9MB model.
A laptop mic in a room with traffic outside comes out close-miked. 5 minutes of audio in 1s, using a 9MB model.
5
Redact: filter PII on the device, in 27 languages.
Mask names, addresses, card numbers as the user types, so PII never reaches your servers. A 12MB model that catches 88.8% of personal data.
Mask names, addresses, card numbers as the user types, so PII never reaches your servers. A 12MB model that catches 88.8% of personal data.
6
Clips: create short videos and highlight clips.
Turn a 10-minute video into a dozen clips in 5s on an iPhone. 10x faster than the Claude Sonnet call we replaced in Detail, at the same quality, with 470x less energy.
Turn a 10-minute video into a dozen clips in 5s on an iPhone. 10x faster than the Claude Sonnet call we replaced in Detail, at the same quality, with 470x less energy.
7
Tongue detects language based on 3 words, from 2MB.
Shapes turns a rough circle into a clean one in under 10ms, from 0.2MB.
Emo suggests emoji in under 2ms.
Small enough to run on every keystroke.
Shapes turns a rough circle into a clean one in under 10ms, from 0.2MB.
Emo suggests emoji in under 2ms.
Small enough to run on every keystroke.
8
Every model is trained on licensed and openly available data.
Every benchmark is public on Hugging Face, against the best available baseline.
Every model spec states the limits in plain words.
Every benchmark is public on Hugging Face, against the best available baseline.
Every model spec states the limits in plain words.
9
Free up to 100k monthly active devices per platform. No limit on how often each person runs a model.
Build the feature you couldn't afford per call.
Build the feature you couldn't afford per call.
10
Little brains in every product.
∗ The launch post: desertant.com/blog/introduci…
∗ Models: huggingface.co/desert-ant-labs
∗ SDK: github.com/Desert-Ant-Labs
∗ The launch post: desertant.com/blog/introduci…
∗ Models: huggingface.co/desert-ant-labs
∗ SDK: github.com/Desert-Ant-Labs