Desert Ant Labs: local, fast models that run on device
Points and comments are a snapshot, not live.
Desert Ant Labs releases 18 small, fast, specialized AI models that run entirely on-device.
The European startup offers models for audio, vision, and text that run in milliseconds on devices, with no per-call cost. Models include Voz (10 min audio transcription in 2 seconds), Clear (9MB audio enhancer), Redact (real-time PII masking in 27 languages), and Tongue (2MB language identifier).
The SDK supports Swift, Kotlin, and JavaScript. Models are free up to 100k monthly active devices. The company argues device-side inference is cheaper and more private than cloud APIs, citing that 40-70% of LLM calls could be handled by small specialized models.
What commenters are saying
Commenters noted that several models appear to be repackaged open-source work (Voz is Parakeet 0.6B v3, Clear is DeepFilterNet 3) with optimized inference code for Apple's Neural Engine. The company confirmed Voz uses Parakeet with custom inference for 300x realtime speed on iPhones, with a next-gen model in development.
Some questioned the business model, since local weights don't require ongoing payment. Others appreciated the free tier and alignment incentives. A common sentiment: useful optimizations, but not new models, and limited to Apple platforms initially.