AI providers
Choose who translates for each feature, from the built-in AI and free services to your own API keys and local models.
A provider is the service that does the translating. Duet Translate can use several at once and lets you choose one for each feature. Everything is on the API Providers page in the settings.
The kinds of provider
| Kind | Examples | You need |
|---|---|---|
| Built-in AI | Duet Translate’s hosted models | A Duet Translate account (plans) |
| Free translation | Google Translate, Microsoft Translator | Nothing |
| Built-in AI providers | OpenAI, Anthropic, Google Gemini, DeepSeek, Mistral and more | An API key from that service |
| Compatible APIs | OpenRouter, SiliconFlow, Volcengine, any OpenAI-compatible endpoint | An API key and sometimes a base URL |
| Translation APIs | DeepL, DeepLX | A key or a server address |
| Local models | Ollama | Ollama running on your computer |
AI models translate more naturally and follow your prompts and glossary. Translation APIs and free services are fast and cheap, but they translate word for word and can’t run AI actions.
Add a provider
- Open API Providers and click Add provider.
- Pick the service. Most need only an API key; Get API key takes you to the page where the service gives you one.
- Choose a model. Fetch available models lists the ones your key can use.
- Click Test connection. If it says Slow, try turning the model’s thinking or reasoning mode down under Advanced options.
Your API keys stay in your browser. Requests go straight from your browser to the service you chose, not through our servers.
Choose a provider for each feature
Under Feature providers, pick the provider for page translation, selection translation, input translation and video subtitles separately. Custom AI actions each choose their own provider in their settings.
A common setup: a fast, cheap provider for whole pages, and a stronger model for selections and AI actions.
Advanced options
- Temperature: lower is more literal and consistent, higher more varied. Leave it empty for the service’s default.
- Reasoning: how much the model thinks before answering, for models that support it. Less is faster and cheaper; translation rarely needs much.
- Headers and Provider options: extra settings some services need. Leave them empty unless the service’s own docs tell you otherwise.
Language detection
Language detection decides how the page’s language is recognised, for auto-translate rules and for picking a voice to read aloud. Basic works offline; LLM asks an AI provider and is more accurate on short or mixed text.