Desert Ant Labs ships 18 free on-device AI models, no API keys
Whisper-beating speech, PII redaction, and language ID that run on an iPhone with no cloud bill — plus a Claude Code update that caps effort per request.
Copy markdown
18 models, on-device, no API keys
Desert Ant Labs released 18 small models — 12 stable, 6 beta — spanning audio, text, and vision that run entirely on-device (iPhone, Android, M-series Macs) through Swift, Kotlin, and JavaScript SDKs. Every model is free up to 100k monthly active devices per SDK, with unlimited inference and no API keys or tokens.
Voz transcribes 4.7x faster than Whisper
Voz turns 10 minutes of audio into text in ~2 seconds on an iPhone and hits 319x realtime on an M3 Ultra — 4.7x faster than OpenAI's Whisper. Companion model Clear (9MB) denoises 5 minutes of audio in ~1 second and reportedly draws 470x less energy than an equivalent Claude Sonnet API call.
2MB language ID, 12MB PII redaction
Tongue identifies 84 languages from three words at 0.933 accuracy in just 2MB; Redact catches 88.8% of personal data across 27 languages at 12MB. Both run locally, so audio and PII never leave the device — useful for compliance-bound apps.
Build this weekend: a fully offline voice pipeline
Wire Voz + Clear + Redact through one SDK and you get a voice recorder that transcribes, denoises, and strips PII entirely offline — zero per-call cost, no cloud round-trip. The free tier covers up to 100k devices, so a real iOS or Android app ships without an inference bill.
Elsewhere: Claude Code 2.1.267 caps effort per request
Claude Code 2.1.267 adds a maxEffortLevel setting to cap reasoning effort across Bedrock, Vertex, and Foundry, plus prompt-cache stability fixes for subagents and --system-prompt sessions. It also fixes Cowork scheduled tasks that were failing at startup under sandboxing.