§ Models
Models on Linux
Dettivo for Linux runs whisper.cpp and parakeet.cpp for speech, llama.cpp for Enhanced and meeting analysis, and sherpa-onnx to tell speakers apart. Every model runs on your machine, on a Vulkan GPU or on the CPU.
Who this is for
Anyone picking a speech model, trying Parakeet Ultra, or clearing disk space.
Speech models
First run recommends the one that transcribes your language best, and dettivo speech providers --language <code> prints the ranking. Meetings transcribe with a Whisper model. The Parakeet models are for dictation, and both engines use your vocabulary.
| Model | Engine | Good for |
|---|---|---|
| large-v3-turbo | whisper.cpp | 99 languages and vocabulary prompts. The default dictation model. |
| small.en | whisper.cpp | English, light enough for a laptop without a GPU. |
| parakeet-v3 | parakeet.cpp | 25 European languages with a time for every word. Fast on a GPU. |
| parakeet-ultra | parakeet.cpp | Moondream's post-train of Parakeet v3. Same languages and size, with a lower word error rate on Moondream's benchmarks. |
| parakeet-v2 | parakeet.cpp | English only. |
Switching to Parakeet Ultra
Terminal
bash
dettivo speech download --provider parakeet --model parakeet-ultra --wait
dettivo speech selection set --provider parakeet --model parakeet-ultraLanguage models
Enhanced and meeting analysis use a Qwen3 model on llama.cpp. dettivo llm download --model <id> --wait fetches one and dettivo config set llm.model <id> selects it.
| Model | For | Size |
|---|---|---|
| qwen3-4b-instruct-2507 | Enhanced and meeting analysis (default) | 2.5 GB |
| qwen3-1.7b | Enhanced, the fast option | 1.1 GB |
| qwen3-4b | Enhanced, the quality option | 2.5 GB |
| qwen3-8b | Meeting analysis | 5.0 GB |
Checked downloads
Every download is checked against a SHA-256 pinned in the catalogue. A file that fails the check goes to quarantine and is never loaded. An interrupted download resumes, and dettivo speech status shows where every model stands, then whether each speech engine is cold, loaded or warm. A warm engine has already run, so the next take pays for no start.
Any speech model download also fetches whisper/silero-vad, the small model the speech gate uses to tell a voice from noise.
Models live under ~/.local/share/dettivo/models/. dettivo speech delete --model <id> removes one, and the selected model needs --force.
Free and open source · GPL-3.0
Dettivo for Linux 0.5.0 is on GitHub.
The release page has the pacman package and its checksum. Bug reports go to the issue tracker. On a Mac? The native Mac app does all of this with Apple Silicon models.