Whistle — 16.9 MB on-device speech recognition

A tiny speech-to-text model (single 16.9 MB file) for phones, wearables, robots, and microcontrollers: transcription, word timestamps, and speech embeddings, all on-device with no GPU or dependencies.

Why it matters

What you could build with it

Does it hold up?

Promising but early: the model card's benchmarks are thorough and self-scored against published Whisper/Moonshine numbers, and the Cactus engine has real adoption — but no third-party device-integration reports for Whistle itself yet.

Built with Whistle

• Cactus Compute releases a 16.9MB speech model for local CPUs (RuntimeWire deep dive)runtimewire · Third-party launch coverage: walks through Whistle's Needle-runtime integration, keyword biasing and word timestamps, with the release video from Cactus's X post.
• Cactus fits speech transcription into a 16.9 MB model (alextech.ai)alextech.ai · Writeup of the Whistle collection on HF: variable-depth decoder, loading alongside Needle for speech-to-tool-call pipelines, 11ms first-token latency claim.
• Whistle: This Tiny 16MB AI Model That Can Hear You (YouTube walkthrough)youtube · Video demo of Whistle's offline transcription on device, 7-language support, and use cases across wearables, robots and agents.
• cactus-compute/needlegithub · 13,419 ★ · The Needle engine repo (13k stars) that Whistle runs on — same binary, same .cact container; Whistle loads via the same needle runtime.

Learn more

official docs →

First spotted on hf: source.

More AI models