Whistle

A 16.9 MB open speech recognizer for seven languages that runs on a plain CPU, with word timestamps and a first token in about 11 ms.

In short. Whistle is a tiny speech-to-text model that fits in 16.9 MB. It transcribes seven European languages, gives a timestamp for every word, and runs on an ordinary processor with no graphics card. It trades some accuracy for size and speed, so it suits phones, wearables and other small devices.

Summary

Whistle is a speech recognition model from Cactus Compute, a San Francisco startup, released on 2 October 2026 under Apache 2.0. The whole model ships as one 16.9 MB file in the company's .cact format and runs on the CPU in a self-contained C++ engine with no dependencies. The Hugging Face repository also holds a 220.6 MB FP32 checkpoint with about 55 million parameters, a count we derived from the file. Whistle transcribes English, German, French, Spanish, Italian, Dutch and Polish, detects the language unless you name it, returns every word with its start time, end time and probability, and can output the encoder's speech embedding, one vector for every 80 milliseconds of audio.

Inside, an 8-block encoder feeds an 8-layer decoder of width 512 through gated cross-attention. The decoder shares its building blocks with Cactus's Needle 3 model, so one engine can load both and turn a spoken request straight into tool calls. Decoding uses five beams, a vocabulary of 8,192 text pieces plus seven language tokens, and optional keyword biasing, which favors names or terms you pass in. Audio must be 16 kHz mono. One pass handles up to 30 seconds and longer clips are windowed. The engine returns an empty transcript when it measures silence.

The developer's blog compares Whistle with only two tiny recognizers, Whisper base (145.3 MB) and Moonshine tiny v2 (41.9 MB), on 10 seconds of audio on an Apple M4 Pro CPU. Whistle reached the first token in 11.1 ms against 73.2 ms for Whisper base and 22.8 ms for Moonshine, and decoded 1,319 tokens per second against 266 and 262. On accuracy, the blog says Whistle is ahead on LibriSpeech test-clean and test-other, SPGISpeech, Earnings-22 and the FLEURS average, while Whisper base is ahead on TED-LIUM, AMI and the MLS average. Whistle's error rates were measured over 86,174 utterances, and the Whisper and Moonshine figures are the ones their authors published. These are the developer's own results and we have not reproduced them.

Cactus publishes the engine as prebuilt binaries for 17 targets, from macOS, Linux, Windows and Android to iOS, watchOS, RISC-V, the browser and WASI. Install it with pip install cactus-compute and call cactus.transcribe on a 16 kHz WAV file. The company has also published guides for fine-tuning Whistle on a new language or your own vocabulary and for porting it to another runtime.

Good for

  • Offline transcription on phones, wearables, Raspberry Pi-class boards and other small devices
  • Voice commands, where a first word after about 11 ms matters more than the last point of accuracy
  • Word-level timestamps and speech embeddings from one small model
  • Commercial products, since the license is Apache 2.0

Watch out for

  • Only seven languages are covered, so there is no Chinese, Japanese, Korean or Arabic
  • The blog compares it only with tiny recognizers, so do not assume it matches larger models such as Parakeet or Cohere Transcribe
  • One pass handles 30 seconds, and longer audio is cut into windows
  • The quoted speeds come from Cactus's own engine and `.cact` file, so other runtimes need a port
  • The accuracy and speed figures are the developer's own

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of Whistle. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
Whistle (FP32 checkpoint) FP32 55M 0.22 GB 0.253 GB Laptop or phone

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Whistle?
Whistle is a tiny speech-to-text model that fits in 16.9 MB. It transcribes seven European languages, gives a timestamp for every word, and runs on an ordinary processor with no graphics card. It trades some accuracy for size and speed, so it suits phones, wearables and other small devices.
Can I use Whistle commercially?
Yes. Whistle is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Whistle open source?
Whistle is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Whistle on my own computer?
Yes. The smallest variant, Whistle (FP32 checkpoint), needs about 0.253 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Whistle released?
Whistle was released by Cactus Compute on October 2, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Needle 3 Cactus Compute 121M 8K Apache 2.0 (Permissive) 0.0835 GB Laptop or phone
Parakeet TDT 0.6B v3 NVIDIA 627M – CC BY 4.0 (Permissive) 2.89 GB Laptop or phone
Granite Speech 5.0 IBM 470M – 2.31B – Apache 2.0 (Permissive) 1.09 GB Laptop or phone
Whisper large-v3-turbo OpenAI 810M – 1.54B – MIT (Permissive) 6 GB Laptop or phone