Screenshot of the first onboarding question, 'Quina llengua llegeixes millor? Which language do you read best?', with buttons for English, Español, Euskara, Galego, Français and DeutschScreenshot of the second question, 'What is your level of Catalan?', listing levels A1 to C2 with a one-line description of each, above Back and Start buttonsThe same level question with A1, 'I can use a few words and simple phrases', selected and the Start button enabledScreenshot of the empty chat, reading 'Escriu un missatge en qualsevol idioma. Et respondré en català.' above a message box, a Send button and a translate toggle
Screenshot of a chat where English questions about the boiling point of water, Antoni Gaudí and Barcelona get short Catalan replies, each with an English gloss and an explanation of key wordsScreenshot of translation mode, where English sentences such as 'did you enjoy the movie?' come back in Catalan, as in 'Vas gaudir la pel·lícula?'Screenshot of the saved exchanges list, with 'can i borrow your phone so i can call claudia?' expanded to show its Catalan translation and a delete buttonScreenshot of the learner profile opened from the header, with English chosen as the reading language, level A1 selected and a Save button

Xat en català

Xat en català (“chat in Catalan”) is a chatbot for learning Catalan. Whatever language you write in, it replies in Catalan, the way a teacher would answer a student, and explains a word or expression when that helps.

On a first visit it asks two questions: which language you read best, and how well you know Catalan, from A1 to C2. Every reply is pitched at that level. Beginners get short sentences, common words and an explanation in their own language; from B1 up, explanations are in Catalan, and fewer as the level rises. A button in the header lets you change your level as you improve.

A toggle turns the chat into a translator, putting whatever you type into Catalan instead of answering it. Replies can be played aloud and saved to a list of exchanges to come back to.

The chat runs on Salamandra 7B Instruct and the speech on Matxa-TTS, both open models from the Barcelona Supercomputing Center’s AINA Kit for building AI in Catalan. The app is built with Next.js and the Vercel AI SDK. The model is quantised and served by llama.cpp on a Google Cloud Run GPU that scales to zero when nobody is using it.

Because the models run on paid cloud GPUs, the demo isn’t public. If you’d like to try it, get in touch. You can also view the source on GitHub.