A 16.9 MB Speech Model for Devices: Cactus Whistle Draws 16K Downloads
Cactus Compute pitches one small model for on-device transcription, word timestamps and audio embeddings across seven languages.
Speech recognition in one small file
Cactus Compute pitches Whistle as a speech model that can run on phones, wearables, robots and other devices without a GPU. The creator says its model is a single 16.9 MB file, yet handles transcription in English, German, French, Spanish, Italian, Dutch and Polish. Its English model card also describes word-level timestamps, audio embeddings and keyword biasing for names or other terms an application needs to recognize.
The model shares an engine and file format with Cactus Compute’s Needle project. According to the creator, an app can transcribe audio locally and pass the transcript into a tool-using workflow through that engine. This is a practical pitch for developers who want speech input on a device rather than a hosted transcription service. The published card includes installation and command-line examples, but the performance and deployment claims are the creator’s, not independent tests by AITrending.
Why it is back in view
Hugging Face dates the Whistle repository’s creation to September 30 and its most recent card update to October 10. In an October 12 trend snapshot, the platform showed 16,434 downloads and 349 likes. These are cumulative repository counters, not distinct users or proof of a sudden surge. The current attention is the news hook; Whistle did not first launch on the date of this story.