Skip to main content
Early accessGnani Evon v3.5 is now available as an API to early access registrants.

When to use this

  • Your users write in an Indian language and expect replies in the same language.
  • You need to translate between English and an Indian language.
  • You need to classify or extract details from messages written in an Indian language.

Supported languages

The model reads and writes 11 languages: The model works best with each language in its own script. Romanised text (an Indian language written in Latin letters, such as “aapka order kab aayega”) and code-mixed text (such as Hinglish) are not in this list, so test them with your own messages before you rely on them.

Set the reply language

Say in the system message which language to reply in, and in which script. Name the language explicitly; do not rely on the model matching the user’s language.
The reply is in choices[0].message.content, in Kannada script.

How it works

The model reads and writes text in tokens. It has no separate language setting: the language of the reply comes from your instructions and the conversation. A clear instruction in the system message is the most reliable way to fix it, because the system message is sent with every request.

Patterns for the system message

Transliteration tips

  • Say which script you want. “Reply in Hindi” can come back in Devanagari or Latin letters. “Reply in Hindi, in Devanagari script” is clearer.
  • Check the script of the reply in your code if it matters, for example before you send text to a text-to-speech voice for one language.
  • Keep identifiers unchanged. Ask the model not to transliterate names, IDs and codes your system needs to match.

Examples

Request body
Request body
The reply’s content is a JSON string such as {"intent": "failed_payment", "language": "hi"}. See JSON output.
Request body
Request body

Voice agents

When the reply is spoken by a text-to-speech voice, speed and length matter more than in a chat window.
  • Turn thinking off with "reasoning_effort": "none". Thinking adds time before the first word of the answer. See Thinking.
  • Ask for short replies. One or two sentences per turn, no lists or markdown. Set max_tokens to match.
  • Stream the reply with "stream": true, and send text to speech as it arrives. See Streaming.
  • Write for the ear. Ask for numbers, amounts and dates in words, in the reply language.
  • Match the languages end to end. Pick the speech-to-text language and the text-to-speech voice to match the reply language. See Speech-to-Text models and Text-to-Speech.

Gotchas

  • Indian scripts can use more tokens. Text in Indian scripts can take more tokens than English for the same meaning, so the same conversation uses more of your context and daily quota. Every response reports the exact count in usage. See Limits.
  • The instruction must be in every request. The API is stateless, so keep the language instruction in the system message you send each time.
  • Personal identifiers are masked in replies, such as phone numbers and UPI IDs, so do not ask the model to repeat them back. See Safety and privacy.

JSON output

Get replies your code can parse.

Limits

Tokens, context size and the daily quota.

Speech-to-Text models

Transcribe what the caller says.

Text-to-Speech

Speak the model’s reply.