Indian LLMs for Voice Agents: What to Know
Jaseem Abbas
- Published

An Indian LLM is a large language model trained with a large share of Indian-language data, often by an Indian lab, to read and write Hindi, Tamil and other Indian languages well. This guide is for the person who has to decide whether a voice agent should run on one. It lists the Indian models that publish open model cards, with facts as each card states them on the date we read it, then explains what a live phone call asks of a language model and how to test it yourself.
Here is the question in a story. Kavya is the CTO of an e-commerce firm in Bengaluru. Her support line takes 6,000 calls a day, most in Kannada, Hindi and English, often in the same sentence. A vendor says, "Our model is built in India, so it is best for Indian languages." Kavya has a better question: does it hold a call in Kannada at 8 pm when the caller mixes in English product names, and does it call the order lookup correctly? Kavya and her firm are made up for this example.
What is an Indian LLM?
An Indian LLM is a large language model trained with a large share of Indian-language data, often by an Indian lab, so that it can read and write Hindi, Tamil and other Indian languages well. The label says where the data and the lab come from. It does not say how the model behaves on a phone call.
A large language model is a program trained on text to predict the next word, and it is the part of a voice agent that decides what to say. Our glossary entry on LLMs explains the idea. People also say "Indic LLM" for a model that covers Indian languages, and "LLM for Indian languages" for the same thing in plain words.
An Indian LLM is only one link in a voice agent. The caller's speech is first turned into text (speech-to-text), the model decides the reply, and a voice reads it out. Some systems merge those steps into one model; see speech-to-speech vs cascaded voice agents. The speech side is separate from the model, and our guide to Hindi speech recognition on calls covers it.
Which Indic LLMs has India built so far?
Labs such as AI4Bharat, BharatGen, Krutrim and Sarvam publish open model cards, and the IndiaAI Mission supports twelve organisations building sovereign foundation models. This is a dated snapshot of what we read, not a ranking.
What does the government say?
The Press Information Bureau, in a release from the Ministry of Electronics and Information Technology dated 25 March 2026, says: "In 2024, Government launched IndiaAI Mission." It adds: "Government is supporting twelve organisations and consortia to develop sovereign foundational model. These include Sarvam AI, Soket AI, Gnani AI, Gan AI, Avatar AI, IIT Bombay Consortium (BharatGen), GenLoop, Zentieq, Intellihealth, Shodh AI, Fractal Analytics Ltd. and Tech Mahindra Maker's Lab."
It also says sovereign models of Sarvam AI, BharatGen and Gnani were launched during the IndiaAI Impact Summit 2026, and that the Sarvam and BharatGen models are available on the AIKosh platform. Read the PIB release on the IndiaAI Mission (read 15 November 2026).
What do the model cards say?
The table lists one model per organisation, in alphabetical order by organisation, with only what each official Hugging Face card says. All cards were read on 15 November 2026. We repeat no "state-of-the-art" claim as fact, and a card's licence line is the card's own.
Organisation: AI4Bharat (IIT Madras)
- Model
- Airavata
- What its card says
- A Hindi instruction-tuned model, fine-tuned from OpenHathi. The paper is on [arXiv](https://arxiv.org/abs/2401.15006).
- Licence as the card gives it
- See the card
Organisation: BharatGen
- Model
- [Param-2-17B-MoE-A2.4B](https://huggingface.co/bharatgenai/Param2-17B-A2.4B-Thinking)
- What its card says
- A mixture-of-experts model with 17 billion parameters and 2.4 billion active per token, 64 experts, about 22.5 trillion training tokens and a context of up to 32,768 tokens. Lists English, Hindi and 21 more Indian languages, and lists tool calling.
- Licence as the card gives it
- Badge reads "BharatGen Non-Commercial"
Organisation: Krutrim (Ola)
- Model
- [Krutrim-2-instruct](https://huggingface.co/krutrim-ai-labs/Krutrim-2-instruct)
- What its card says
- A 12 billion parameter dense model built on Mistral-NeMo 12B, with a context of up to 128K tokens. The card itself notes that many Indic benchmarks are translations of English ones.
- Licence as the card gives it
- Krutrim Community License Agreement
Organisation: Sarvam
- Model
- [Sarvam-105B](https://huggingface.co/sarvamai/sarvam-105b)
- What its card says
- A mixture-of-experts model with 10.3 billion active parameters, listing 22 Indian languages.
- Licence as the card gives it
- Apache 2.0
Other organisations publish cards too. For example, BharatGen also has a 2.9 billion parameter Hindi and English model, and Sarvam publishes smaller models as well. For Sarvam's APIs for speech and voice, read the Sarvam AI API guide. We show one model per organisation so that no one gets more space than another. Gnani AI appears in the PIB list above. We did not read a model card for it, so it has no row.
What does a voice agent need from a language model?
A voice agent needs fast first words, replies in the script the voice reads, sense from mixed Hindi and English, reliable tool calls and a licence that allows commercial use. A general benchmark score covers none of these directly.
Need: Fast first words
- Why it matters on a call
- The caller hears silence until the model starts to reply
- How to check it
- Time the first word of the reply on your own prompts, many times, and look at the slow ones
Need: Replies in the right script
- Why it matters on a call
- The voice reads what the model writes. A reply in the wrong script, or with stray symbols, is read out loud or read wrongly
- How to check it
- Read 50 replies aloud in each language you serve
Need: Sense from mixed speech
- Why it matters on a call
- Callers say "mera order status kya hai" in one breath. Our guide to [Hinglish voice agents](/blog/hinglish-voice-agents/) explains why this is normal
- How to check it
- Test with real mixed sentences, not only clean Hindi. See [code-switching](/glossary/code-switching/)
Need: Reliable tool calls
- Why it matters on a call
- The model must call the order lookup with the right order number. See [function calling](/glossary/function-calling/)
- How to check it
- Score tool calls separately from the reply text
Need: A licence that fits
- Why it matters on a call
- Some cards allow commercial use and some do not
- How to check it
- Read the licence on the card itself
The speech-to-text step and the voice carry their own weight. If the transcript is wrong, the best model answers the wrong question. See our guides to call transcription in Indian languages and Hindi speech recognition on calls.
Why does model size matter on a phone call?
Every word waits on the model, so a mixture-of-experts model with few active parameters can reply faster than its total size suggests. A dense model uses all its parameters for every word. A mixture-of-experts (MoE) model holds many "experts" and uses only a few for each word, so its active size is what counts for speed.
The cards show the difference. Krutrim-2-instruct is a 12 billion parameter dense model, per its card. Param-2-17B-MoE-A2.4B holds 17 billion parameters but uses 2.4 billion per token, per its card. Sarvam-105B uses 10.3 billion active parameters, per its card. We publish no latency number, because speed also depends on the hardware, the length of the prompt and how the model is served.
Read our guide to voice agent latency for where the model's share sits among the other delays on a call.
Why does tokenisation matter for Indian languages?
Indian scripts can take more tokens per word, which adds time and cost to every reply. A token is a piece of a word that the model reads and writes. A model that splits a Hindi word into many small pieces needs more steps to say it.
The measure is "fertility", the number of tokens per word. The card for Sarvam-1 reports "fertility rates of 1.4-2.1" across its languages (read 15 November 2026). Lower is better. We quote one card, because we did not read comparable figures for the others, and a number from one card cannot be set against a number from another without the same method.
What do public Indic benchmarks show?
Public benchmarks show that models know Indian languages unevenly, and none of them tests a phone call. Two original studies are worth knowing.
- MILU from AI4Bharat and others covers "8 domains and 41 subjects across 11 Indic languages" and evaluated "over 42 LLMs". It reports "GPT-4o achieving the highest average accuracy at 74 percent", that "open multilingual models outperform language-specific fine-tuned models", and that models do worse in "culturally relevant areas like Arts and Humanities, Law and Governance". Read the MILU paper on arXiv (read 15 November 2026).
- IndicGenBench from Google Research covers "29 Indic languages covering 13 scripts and 4 language families" with generation tasks such as cross-lingual summarisation, translation and question answering. Read the IndicGenBench paper on arXiv (read 15 November 2026).
A benchmark tells you that a model knows Hindi. It does not tell you that the model holds a conversation with an interrupting caller, or that it calls your order API correctly. For that you test.
How do you test an LLM for Hindi and Tamil calls?
Run your own call scripts, in the scripts and mixes your callers use, by voice and by text, and score tool calls and wrong facts separately. A model that wins a leaderboard can still fail your top ten call types.
- Collect 50 to 100 real call openings from your own line, in each language. If you have none, write them with native speakers.
- Cover the mixes. Pure Hindi, Hinglish, Tamil with English words, numbers, names and addresses.
- Run the same set through each model by text first, then by voice with your speech-to-text and voice in place.
- Score three things apart. Did the reply make sense? Was a fact wrong? Was the tool called with the right values? A wrong fact is a hallucination, and our glossary entry on AI hallucination explains it. If your agent answers from your documents, see retrieval-augmented generation.
- Listen to the slow ones. Look at the slowest 10% of first words, not the average.
- Repeat after any change to the prompt, the model or the voice.
Our guide to how to test a voice agent goes through the full test plan, and the voice agent prompt guide shows how a prompt changes the results.
Should you pick an Indian model or a global one?
Pick on your own test results, the licence and where the model can run, not on where it was built. A model built in India may be a good fit for your Tamil calls, or it may not. A global model may handle Hindi well, or it may not. The MILU paper found open multilingual models outperformed language-specific fine-tuned ones, which is one more reason to test and not assume.
Four questions decide it.
- Did it pass your tests in the languages and mixes your callers use?
- Does the licence allow your use, commercial use included? Cards differ.
- Where can it run? Some firms need a model that runs in India, or on their own servers.
- What does a reply cost and how fast is it, on your own traffic?
How does Bananaflow fit?
Bananaflow is an AI voice agent platform. Each agent picks its own speech-to-text, language model and voice, and any one of them can be changed without rebuilding the agent. The platform handles English (India) and 10 Indian languages, each heard and spoken in its own script. See the languages page.
The agent follows the caller's language and switches mid-call, and Hinglish is written down as spoken. You can test any agent in the browser by voice call or text chat before it takes real calls. Call data is not used to train models.
What should you remember about Indian LLMs for voice agents?
An Indian LLM is a model trained with a large share of Indian-language data. A phone call needs fast first words, the right script, sense from mixed speech, reliable tool calls and a licence that fits. Compare active size for speed, read each card as a record and not a verdict, and test with your own calls. Pick on your results.
FAQ
Frequently asked questions
Which is the best Indian LLM?
There is no single answer. The best model for you is the one that passes your own call tests in your languages, with a licence that fits and a speed your callers accept. Public benchmarks show wide gaps: the MILU study found the top model it tested reached 74% average accuracy across 11 Indian languages, and every model did worse on local law and culture.
Are Indian LLMs open source?
Some are open, and the licences differ. As read on 15 November 2026, the cards for Sarvam's models list Apache 2.0, Krutrim-2-instruct uses the Krutrim Community License Agreement and BharatGen's Param-2-17B card carries a BharatGen Non-Commercial badge. Read the licence on each card before you use a model in a product.
What is the IndiaAI Mission?
It is a Government of India programme launched in 2024. According to the Press Information Bureau, on 25 March 2026 the Government was supporting twelve organisations and consortia to develop sovereign foundation models. It also reported 38,231 GPUs onboarded and made the Sarvam and BharatGen models available on the AIKosh platform.
Do voice agents need an Indian LLM to speak Hindi?
No. A voice agent needs a language model that passes your Hindi call tests, plus speech-to-text and a voice that handle Hindi well. The model may be Indian or global. Test the full chain on real call openings, including Hinglish, and pick on the results.
What is token fertility?
Token fertility is the number of tokens a model uses per word. Indian scripts can take more tokens per word, so a higher fertility means more steps, more time and more cost on every reply. The Sarvam-1 model card reports fertility rates of 1.4 to 2.1 across its languages.
Does a bigger model mean a slower voice agent?
Not always. A dense model uses all its parameters for every word, while a mixture-of-experts model uses only some, so its active size is what to compare. For example, Param-2-17B-MoE-A2.4B holds 17 billion parameters and uses 2.4 billion per token, per its card. Speed also depends on hardware and how the model is served, so time it yourself.




