Whistle is a speech-to-text model from Cactus Compute aimed at mobiles, wearables, robots, smart home, automotive and microcontrollers. Cactus describes the whole model as a single 16.9 MB file that runs on a CPU engine with no dependencies and no GPU. It transcribes 16 kHz mono audio, up to 30 seconds in one pass, in English, German, French, Spanish, Italian, Dutch and Polish; the language is detected unless you name it, and silence returns an empty transcript rather than an invented sentence. Beyond transcription it returns word timestamps (start, end and probability for each word, aligned from the decoder's own attention) and a speech embedding, the encoder output with one row per 80 ms frame, for matching and retrieval without decoding a transcript. Keyword biasing favours names, places and product words during the search. Architecturally it is a log-mel front end and a convolutional stem feeding an audio encoder, read by a Needle-shaped decoder through gated cross attention at every layer; the decoder is laddered so any depth from 2 layers up can be picked at load time with --audio-depth. It reuses the .cact container, quantisation, SIMD kernels and KV cache of Cactus's Needle engine, so a device that already runs Needle can run Whistle in the same binary, or load both models and turn a clip into tool calls in one call. Precision is 2 to 4 bit. The card reports word error rates on full test splits scored with the Whisper normalizers, with Whisper and Moonshine figures taken as published by their authors, and a speed test on an Apple M4 Pro with 10 s of audio and Whistle at 5 beams, each model on its official runtime. Cactus states that no test audio appears in Whistle's training or validation data. Install with pip install cactus-needle; the weights are Apache-2.0.