Cactus releases Whistle speech model at 16.9 MB running on CPU

Cactus launched Whistle on October 2, 2026, an open-source speech recognition model weighing 16.9 MB that runs entirely on CPU with no external dependencies. It transcribes seven languages, reaches first token in 11.1 ms on an Apple M4 Pro and decodes at 1,319 tokens/s — about five times faster than Whisper base.

• Model weighs 16.9 MB — under 1/8 of Whisper base (145.3 MB)
• First token in 11.1 ms on M4 Pro vs 73.2 ms for Whisper base
• Decoding speed of 1,319 tokens/s, 5x faster than Whisper base
• Supports 7 languages: English, German, French, Spanish, Italian, Dutch, Polish