3.12.1
October 18, 2026VivaDicta 3.12.1 brings new AI models, keeps your transcription language where you set it, and gets the keyboard back to more apps after dictating.
AI model updates
- Gemini 4 Argon, DeepSeek V4.1 Flash, Qwen 3.8 Flash and MiMo V2.6 Flash join the AI enhancement model pickers
- DeepSeek V4.1 Flash is available through OpenCode Zen; Qwen 3.8 Flash through Zen and Go; MiMo V2.6 Flash through Go. DeepSeek V4.1 Flash is the new OpenCode Zen default for new selections
- Claude Sonnet 5.5 and Haiku 5.5 are available with an Anthropic API key. GPT-6.1 Sol is also available with ChatGPT sign-in
Fixes and compatibility
- Gemini sign-in applies the same model-specific thinking level as the API-key path
- OpenCode Zen free promotions restricted to OpenCode itself are no longer offered. A mode still set to one now says so plainly instead of failing with a server error, and is never switched to a paid model behind your back
- Groq vocabulary hints no longer bias transcription toward English; automatic language detection sends no vocabulary prompt
- The language selected for a shared file applies only to that import instead of sticking to later dictations
- AI enhancement preserves the transcript language unless the chosen preset asks for a different language
- ChatGPT sign-in no longer offers GPT-5.4 and GPT-5.4 mini, which the endpoint does not serve
- Soniox real-time transcription uses its new authentication method
Return to your apps
- More return links for host apps, including OneNote, Ivory, Tesla, Taimi, Unwatched, Nagram, Cherry Studio and Heeler. Apps without a verified return URL use the manual return path
- Orca joins Safari and Claude under Skip Apps That Lose Your Place in Advanced settings: swipe back after dictating and you land in the same chat instead of Orca's home screen