sarvam-105b-conversations) is the conversational variant of Sarvam 105B, tuned for real-time dialogue and voice agents. It handles native-script, romanized and code-mixed input in 10 Indian languages and English.
Quick config
In the dashboard, open the agent’s Intelligence tab, pick Sarvam as the provider and Sarvam 105B Chat as the model. Over the API:Supported models
Key settings
reasoning_effort and verbosity don’t apply to this model.
Things to plan for
- 32K context window. The limit covers the whole request: system prompt, conversation history, tool definitions and the reply (
max_tokens). A request over the limit fails and the call ends, so keep long prompts and knowledge-heavy nodes well under it. - Ending calls. With the
end_calltool enabled, the model tends to end the call as soon as the task in the prompt looks done, sometimes before it confirms the details back to the caller. Test how your calls end before going live. - Graph agents. Routing between nodes runs on Bolna’s routing model; only the replies come from Sarvam. A node can also switch to Sarvam through its own
llm_config. - Prompt caching. Sarvam caches a repeated prompt prefix automatically, which speeds up later turns of a call. No configuration is needed.
FAQ
Which languages does it support?
Which languages does it support?
Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu and English, in native script, romanized or code-mixed. Pair it with Sarvam speech-to-text and Sarvam text-to-speech for an all-Sarvam agent.
Related
- Intelligence Tab: configure the LLM in the dashboard
- Sarvam text-to-speech and Sarvam speech-to-text
- Custom LLM inference: bring your own OpenAI-compatible endpoint
- Prompting Guide: write effective prompts for voice