Desert Ant Labs Bets 18 Tiny Models Beat One Big API
Paul Veugen spent five years trying to put on-device features into Detail, a video editing app, and eventually concluded the hard part was that nobody makes the models he needed. So on Monday he launched Desert Ant Labs with 18 of them, 12 stable and 6 in beta, all built to run on the device in your hand.
They are aggressively unglamorous and that is the point. Voz does transcription and chews through ten minutes of audio in two seconds on an iPhone, which they claim is 4.7x faster than Whisper. Clear cleans up five minutes of audio in one second from a 9MB file. Redact masks personal data across 27 languages. Tongue identifies which language you are speaking from three words using 2MB. Across the suite they quote real-time factors of 298 to 345x on current hardware. iPhone, MacBook, Android, and browser via WebAssembly, with SDKs for Swift, Kotlin and JavaScript. Free up to 100,000 monthly active devices.
The thesis is stated plainly: the best path to efficient intelligence starts on-device. Not because privacy is nice, though it is, but because inference that costs nothing and returns in milliseconds changes what you are willing to build. If detecting language costs an API call and 300ms, you do it once. If it costs 2MB of RAM and no money, you do it on every keystroke.
This is the same pressure showing up everywhere the agent loop meets a bill. oMLX made local coding agents actually usable on a Mac (https://clauday.com/article/99ad10c7-4e6f-4d83-9545-cd8c8e25d969), and somebody streamed a 2.8T model off four SSDs at one token per second just to prove the ceiling is a budget, not a law (https://clauday.com/article/ca7fd597-8ae5-40fc-997a-97b01da6bdbd). Desert Ant is the other end of that: stop shrinking the frontier model, ship a drawer of small ones that never leave the phone. https://desertant.com/blog/introducing-desert-ant-labs/
← Back to all articles
They are aggressively unglamorous and that is the point. Voz does transcription and chews through ten minutes of audio in two seconds on an iPhone, which they claim is 4.7x faster than Whisper. Clear cleans up five minutes of audio in one second from a 9MB file. Redact masks personal data across 27 languages. Tongue identifies which language you are speaking from three words using 2MB. Across the suite they quote real-time factors of 298 to 345x on current hardware. iPhone, MacBook, Android, and browser via WebAssembly, with SDKs for Swift, Kotlin and JavaScript. Free up to 100,000 monthly active devices.
The thesis is stated plainly: the best path to efficient intelligence starts on-device. Not because privacy is nice, though it is, but because inference that costs nothing and returns in milliseconds changes what you are willing to build. If detecting language costs an API call and 300ms, you do it once. If it costs 2MB of RAM and no money, you do it on every keystroke.
This is the same pressure showing up everywhere the agent loop meets a bill. oMLX made local coding agents actually usable on a Mac (https://clauday.com/article/99ad10c7-4e6f-4d83-9545-cd8c8e25d969), and somebody streamed a 2.8T model off four SSDs at one token per second just to prove the ceiling is a budget, not a law (https://clauday.com/article/ca7fd597-8ae5-40fc-997a-97b01da6bdbd). Desert Ant is the other end of that: stop shrinking the frontier model, ship a drawer of small ones that never leave the phone. https://desertant.com/blog/introducing-desert-ant-labs/
Comments