Mistral moves to Voxtral 3 and gets a cleaned-up model list (#754).
- Voxtral 3: Voxtral Mini Transcribe 3 is the new default, and Realtime 3 replaces Realtime 2. Saved Voxtral 2 selections move to their successors.
- Model list: Refresh models shows one named entry per model instead of every alias Mistral lists, for example "Mistral Large 4". Code, Labs and speech models no longer appear for text processing. Without a refresh, the current lineup is ready: Mini and Small Transcribe 3, Realtime 3, Mistral Large 4, Medium 3.5, Small 4 and Ministral 3 14B, 8B and 3B.
- Dictionary: your dictionary terms now reach Mistral, including names with spaces such as "Saoirse Ronan".
- Languages: Transcribe 3 and Realtime 3 offer 58 languages instead of 13, and the detected language is reported.
- Live text: Realtime 3 shows the text while you speak, and the final text comes from Mini Transcribe 3. In our tests it misheard fewer words and was done about half a second after you stop. Files and recordings without live text also use Mini Transcribe 3. Live text bills both models.
- Reasoning effort: Mistral Large 4, Medium 3.5 and Small 4 get a reasoning effort setting that starts at the fastest level, so a correction with Large 4 takes a few seconds instead of about ten.
- Mistral's error messages appear as text instead of raw JSON.
Requires plugin host 1.1.2, which every TypeWhisper for Windows from 1.1.0 provides.
Source commit: 3560406