Opera 0.9.14 is the biggest update of the series so far. It brings a full French translation, models that run on your device, more servers to talk to, and a redesigned look.
Opera in French
Every string in the app now goes through typed translation catalogs, and the app follows your device language. French is the first translation, and it was polished over the cycle.
Models on your device
- Run models on the device with LiteRT, with model avatars and short descriptions.
- Browse and download models from Hugging Face.
- An on-device model sheet lets you manage them, and it is navigated like the settings pages.
- Opera suggests models from the same family and adds capability icons to model rows.
- The model selector shows token usage in a progress bar, with a better estimate of the model window.
More servers, and failover
- Add any OpenAI-compatible server. API keys are stored securely.
- Each server has its own subpage, and Ollama can be tuned per server.
- When a server fails, Opera fails over automatically to the next one.
- Provider errors are visible in a new message details sheet.
Tools ask first
A tool consent gate asks before a tool runs, through a shared modal. You can also mention models and add MCP servers from a sheet, and desktop tool execution is more reliable.
A new look
- Onboarding is a single page with a profile card, an animated wordmark and cycling phrases.
- The home screen has an interactive butterfly cluster that animates into incognito mode.
- Drawers have floating back buttons, settings pages are wrapped in content cards, and icon labels and tooltips appear across the interface.
- Code and image preview sheets open from the conversation.
- Android gets predictive back.
Other improvements
- Queued replies show as busy.
- Cloud sync payloads are lighter and backups feel smoother.
- Tool schemas are sent without indentation, which saves tokens.
- Selection handles are oriented correctly, and the detector model is quantized.
Opera is open source. The full list of changes is on GitHub.