Cactus Whistle: a 16.9MB speech-to-text model that runs on CPU
On-device voice with no GPU and no API bill — plus a leaked Gemini mode that wants your whole Mac, and a DeepSeek desktop harness.
Copy markdown
16.9MB, on the CPU, no dependencies
Cactus's Whistle hits the first token in 11ms and decodes 1,319 tokens/sec — roughly 8.6x smaller and 6x faster than Whisper base, with lower word-error rates on LibriSpeech, SPGISpeech and Earnings-22. `pip install cactus-needle` then `needle.transcribe("clip.wav")`; weights are on Hugging Face and the engine deploys to 17 targets from microcontrollers to the browser. On-device dictation with zero cloud round-trip is now a weekend project.
Report: Gemini eyes full Mac access
TestingCatalog surfaced an unreleased "Additional sandbox options" in Gemini Desktop that would reportedly let Gemini read, create, modify or delete files anywhere on a Mac and drive apps like Mail and Safari without asking first. Google hasn't confirmed it, and buying, accounts and legal terms would still need approval — but it's a reminder to scope exactly what you let an agent touch.
DeepSeek Harness v0.2 ships desktop installers
The DeepSeek Harness agent added desktop builds for macOS, Windows and Linux with plugin management and workspaces — a local-first way to run DeepSeek models for coding and everyday work without living in a terminal.