Desert Ant Labs launches 18 on-device models that beat cloud APIs on speed and cost

Desert Ant Labs, a European AI lab, has launched with 18 specialized on-device models for audio, vision, and text. The pitch: small models that run in milliseconds on consumer hardware, with zero per-inference cost, so developers can embed intelligence into every interaction without watching token bills climb.

The models are accessible via a single SDK for Swift, Kotlin, and JavaScript. Notable entries include:

• Voz: transcribes 10 minutes of audio in two seconds on an iPhone, claiming 4.7x faster than Whisper, with word-level timestamps.
• Clear: a 9MB model that enhances a five-minute recording to studio quality in one second.
• Redact: masks PII in real time across 27 languages, catching 88.8% of personal data in a 12MB model—close to the 91.1% of a 2.3GB GLiNER-PII model.
• Tongue: identifies 84 languages from three words with a 2MB model, scoring 0.933 accuracy versus 0.887 for a 293MB detector.

The company's origin story is grounded in their video app Detail, where cloud API costs grew with popularity. They trained their own models to replace cloud services like Dolby and Claude Sonnet—for instance, their Clips model turns a 10-minute video into a dozen clips in 5 seconds, claiming 10x faster and 470x less energy than Sonnet with equal quality.

Desert Ant Labs argues that much of the industry's AI workload doesn't need frontier models. They cite NVIDIA research suggesting 40-70% of agent calls could be handled by small, specialized models. They also point to the compute already in users' hands: over a billion phones shipped annually with capable chips, versus $450 billion spent on data centers this year.

The company's roadmap includes a "cortex" layer that routes tasks to the appropriate model—local first, cloud only when necessary. Their models are free up to 100k monthly active devices, with no tokens or logins required.