tl;dr: Cactus released Whistle, a 16.9 MB on-device speech recognition model that transcribes seven languages, runs on CPU with no dependencies, and shares its C++ engine with their Needle LLM so one binary can turn audio directly into tool calls. Benchmarks claim 11ms time-to-first-token and 1,319 tokens/sec decode on an M4 Pro, beating Whisper base on several WER tests at ~1/8th the size, though Whisper leads on TED-LIUM, AMI, and MLS. Weights are on Hugging Face with prebuilt binaries for 17 targets including mobile, embedded, and WASM.