Google Gemini
chat, vision, speech synthesis, image and video generation
Technology
AI models in Vio
Vio is not tied to a single vendor. Each task has its own model, with fallbacks configured on the routes that support them.
A request goes to the task's primary model. If the provider is unavailable, returns an error or times out, the gateway retries on the next model by priority - usually without you noticing.
Different tasks need different things. Parsing a phrase into a date and attendees is precise work that benefits from reasoning, so a model with thinking enabled sits there. A live conversation matters more when it is fast, so the primary model in chat is a fast one.
Text and reasoning
DeepSeek V4 Pro runs in thinking mode: it is used where parsing accuracy matters more than response speed.
| Task | Primary model | Fallback / options |
|---|---|---|
| AI chat | Gemini 3.6 Flash | GPT-5.6 TerraDeepSeek V4 FlashDeepSeek V4 ProGPT-5.6 Luna |
| Answers from your material (RAG) | DeepSeek V4 Pro | Gemini 3.6 FlashGPT-5.6 TerraDeepSeek V4 FlashGPT-5.6 Luna |
| Note summarisation | DeepSeek V4 Pro | Gemini 3.5 Flash LiteDeepSeek V4 FlashGPT-5.6 Luna |
| Calendar event extraction | DeepSeek V4 Pro | Gemini 3.5 Flash LiteDeepSeek V4 FlashGPT-5.6 Luna |
| Task extraction from text | DeepSeek V4 Pro | Gemini 3.5 Flash LiteDeepSeek V4 FlashGPT-5.6 Luna |
| Knowledge base query | DeepSeek V4 Pro | Gemini 3.5 Flash LiteDeepSeek V4 FlashGPT-5.6 Luna |
| Math expression parsing | DeepSeek V4 Pro | Gemini 3.5 Flash LiteDeepSeek V4 FlashGPT-5.6 Luna |
Semantic search
Embeddings let search understand meaning rather than word matches. That is what makes «where did I write this down» work.
| Task | Primary model | Fallback / options |
|---|---|---|
| Content embeddings | OpenAI text-embedding-3-large | Gemini Embedding 001 |
Voice
Recognition works both in the bot and on the site: a voice message becomes the text of a note draft.
| Task | Primary model | Fallback / options |
|---|---|---|
| Speech recognition | GPT-4o mini Transcribe | - |
| Text to speech | Gemini 3.1 Flash TTS | GPT-4o mini TTS |
Images and video
In the image studio the model is chosen manually: Lite is cheaper, Pro is higher quality. The cost is shown before the generation starts.
| Task | Primary model | Fallback / options |
|---|---|---|
| Questions about an image | Gemini 3.6 Flash | GPT-5.6 Terra |
| Image generation | Nano Banana 2 (Gemini 3.1 Flash Image) | Nano Banana 2 Lite (3.1 Flash Lite Image)Nano Banana Pro (3 Pro Image) |
| Video generation | Gemini Omni Flash (preview) | - |
Providers
Vio currently works with three model providers.
DeepSeek
text parsing, answers from your material
OpenAI
embeddings, speech recognition, fallback models
What matters here
- A model only reaches the material you allowed in the chat settings: sources and tools are switched on separately.
- Every answer carries links to the records it was built from - you can open and verify them.
- Buttons without the AI mark («To task», «To calendar») work without calling any model and are available at any quota.
- Your AI usage is visible on the same page where you spend it.
The set of models can change: we switch to more accurate or faster ones as they appear. This page is updated together with the gateway configuration.
Still have a question?
Message on Telegram