Whistle: an open speech-to-text model that fits in 16.9 MB
Cactus's CPU model hits the first word in 11 ms across seven languages — plus a free library of ready-made OAuth skills for your coding agent.
Copy markdown
Speech-to-text that fits in a single 16.9 MB file
Cactus's Whistle runs entirely on-device on the same CPU engine as its Needle LLM, transcribing seven languages with word timestamps and hitting the first token in 11 ms — 6.6x faster than the 145 MB Whisper base, which it also beats on LibriSpeech, SPGISpeech and Earnings-22. Open weights on Hugging Face plus prebuilt binaries for 17 platforms (iOS, Android, the browser, even RISC-V) mean you can drop dictation or voice commands into a mobile or wearable app with zero cloud bill and no network round-trip.
A free OAuth skill library for Claude Code and Cursor
Unified.to published a catalog of free SKILL.md files that teach any coding agent how to handle OAuth and wire up third-party APIs — CRM, accounting, ticketing — alongside typed task objects that track tokens, file changes and human checkpoints. Instead of hand-rolling each integration's auth dance yourself, you point the agent at the matching skill and let it follow the steps.