India's Own AI Models in 2026: Which Ones You Can Actually Use Today, and How
Short answer: as of November 2026, the Indian-built AI you can actually use today is mostly Sarvam. Its Indus chat app is live on Android, iOS and the web, its API gives every verified user ₹100 of free credits, and its 30B and 105B models are open weights under the Apache 2.0 licence. BharatGen Param2 is downloadable too, but it is a research checkpoint under a non-commercial licence and needs a large GPU. Most other names from the headlines are either previews, announcements without public access, or, in Krutrim's case, offline.
India's AI story in 2026 has had plenty of headlines: the India AI Impact Summit in February, 12 government-funded model projects, and claims of models that beat ChatGPT. What the headlines rarely answer is the question most people actually type into Google: can I use it right now, and how?
This guide answers that for every major Indian model. We checked each one against its own model card, pricing page or launch coverage, and we list every source at the end.
Indian AI models at a glance (November 2026)
| Model | Built by | What it is | Can you use it today? | Licence |
|---|---|---|---|---|
| Sarvam 105B | Sarvam AI, Bengaluru | Mixture-of-Experts text model, 106B total, about 10B active, 128K context | Yes. Indus app, Sarvam API, open weights | Apache 2.0 |
| Sarvam 30B | Sarvam AI | Smaller Mixture-of-Experts model, 32B total, about 2.4B active | Yes. Open weights on Hugging Face and AIKosh | Apache 2.0 |
| Sarvam speech and vision (Saaras, Bulbul, Sarvam Vision) | Sarvam AI | Speech-to-text, text-to-speech and Indic document OCR | Yes. Paid API with free starting credits | Commercial API |
| Param2 17B | BharatGen (IIT Bombay-led consortium) | Mixture-of-Experts, 17B total, 2.4B active, 22 Indian languages plus English | Partly. Weights on Hugging Face, no public chat app | BharatGen non-commercial |
| Project Indus 8B | Tech Mahindra | Hindi-first education model | Not publicly. Announced February 2026, no public access route stated | Not stated |
| Inya VoiceOS | Gnani.ai | 5B voice-to-voice model for 15+ Indian languages | Preview only | Not stated |
| Pragna 1B | Soket AI | Small 1.25B model, English plus six Indian languages | Yes, for developers. Open weights | Permissive, check model card |
| Kruti assistant | Ola Krutrim | Agentic AI assistant | No. Offline since April 2026 | Not applicable |
The rest of this post explains each row, how to try the ones that are live, and how good they really are.
Sarvam: the one you can use today
Sarvam AI was founded in Bengaluru in August 2023 by Vivek Raghavan and Pratyush Kumar, both previously linked to AI4Bharat at IIT Madras. In April 2025 the government selected it under the IndiaAI Mission to build an indigenous foundation model, with subsidised GPU compute for training.
At the India AI Impact Summit in February 2026, Sarvam announced two models trained from scratch in India. Unlike its earlier Sarvam-M, which was built on Mistral Small, these were pre-trained from the ground up.
| Sarvam 105B | Sarvam 30B | |
|---|---|---|
| Architecture | Mixture-of-Experts | Mixture-of-Experts |
| Total parameters | 106B | 32B |
| Active parameters per token | About 10B | About 2.4B |
| Context window | 128K tokens | About 65K tokens |
| Modes | Reasoning and non-reasoning | Reasoning and non-reasoning |
| Licence | Apache 2.0 | Apache 2.0 |
Apache 2.0 matters. It means you can download the weights, run them on your own servers and build commercial products on them, which is not true of every "open" Indian model.
How to try Sarvam in three ways
- The Indus app (easiest). Indus launched on 20 February 2026 on the web, the Google Play Store and the Apple App Store. It is the chat interface for Sarvam 105B. You can type or speak, and it can answer in text or audio. You sign in with a phone number, Google, Microsoft or Apple ID. At launch Sarvam warned of a possible waitlist because of limited compute.
- The Sarvam API (for developers). Sign up on Sarvam's developer platform. Every verified user gets ₹100 of free credits that work across all its APIs and do not expire. The chat model sarvam-105b costs ₹15 per million input tokens and ₹60 per million output tokens, with cached input at ₹5, all excluding GST.
- Open weights (for self-hosting). Download
sarvamai/sarvam-30borsarvamai/sarvam-105bfrom Hugging Face, or get them from AIKosh, the government's AI dataset and model platform. The 30B model has community quantised versions for llama.cpp, Ollama and LM Studio, according to its model card, so a strong desktop can run it.
Sarvam's other APIs
Sarvam's paid API also covers the tasks where Indian languages matter most:
- Speech-to-text: ₹0.49 per minute on pay as you go.
- Text-to-speech: Bulbul v3 at ₹3 per 1,000 characters, and a faster Bulbul v4 Flash at ₹1.
- Translation: ₹2 per 1,000 characters, formal or colloquial.
- Sarvam Vision: a 3B document model for OCR in Indian scripts. Sarvam reports 84.3% on olmOCR-Bench, which it says beats Gemini 3 Pro and ChatGPT on multi-script documents.
For an app that needs Tamil voice input or reads scanned Marathi forms, this is the most practical Indian option today.
How good is Sarvam 105B, honestly?
Independent testing gives a mixed picture, and it is worth knowing before you build on it.
- General intelligence: Artificial Analysis scores Sarvam 105B at 18 on its Intelligence Index in reasoning mode. Comparable open models score higher: GLM-4.5-Air 23, Mistral Small 4 at 27 and gpt-oss-120B at 33. Sarvam 30B scores 12, against 22 for Qwen3 30B A3B.
- Agentic tasks: better than its general score suggests. Sarvam 105B scores 25 on the Artificial Analysis agentic index, ahead of GLM-4.5-Air at 21.
- Hallucination: the weak spot. On AA-Omniscience it scores -60, which Artificial Analysis attributes to high hallucination rates and to the model attempting questions instead of saying it does not know.
- Indian knowledge: on MILU, a benchmark of Indian knowledge, Sarvam reports 76.8 for its 30B model, against 82.6 for Qwen3-30B-Thinking and 73.7 for gpt-oss-20b.
The honest summary: Sarvam is not a frontier model. Its real strengths are voice in Indian languages, Indian scripts and documents, a tokenizer that Sarvam says needs fewer tokens for Indian text, and a permissive licence. For hard reasoning or coding, global models are still ahead.
BharatGen Param2: open, but for builders
BharatGen is a government-backed initiative that began as an IIT Bombay-led consortium. It has the largest allocation of any IndiaAI foundation model project: ₹1,058.52 crore, according to a MeitY reply in the Lok Sabha reported by MediaNama in April 2026. Its plan spans models from 2B to 1 trillion parameters.
Its first big public release, Param2, launched at the India AI Impact Summit. The model card for bharatgenai/Param2-17B-A2.4B-Thinking on Hugging Face gives the details:
- Size: 17B total parameters, 2.4B active per token, using 64 experts.
- Training: about 22.5 trillion tokens.
- Languages: English, Hindi and 21 more Indian languages, including Assamese, Bodo, Dogri, Konkani, Kashmiri, Maithili, Manipuri, Sanskrit, Santali and Sindhi, which few global models cover.
- Context: 32,768 tokens, with a 128K variant described as coming soon.
- Hardware: about 34 GB of GPU memory to load in bfloat16.
- Licence: the BharatGen non-commercial licence.
Two points matter for anyone thinking of using it. First, the model card calls it an early post-training checkpoint that needs further fine-tuning and safety alignment before production use. Second, the non-commercial licence means you cannot simply ship it inside a paid product. Check the licence file before you build anything on it.
There is no public BharatGen chat app for consumers. BharatGen has worked on applications with partners, including MahaGPT for the Maharashtra government and Gyan Bharatam for the Ministry of Culture, but those are not general chatbots you can download.
How to try it: if you have a GPU with about 40 GB of memory, load it with Hugging Face Transformers version 4.52.3 and trust_remote_code=True, or serve it with vLLM or SGLang. If you do not, a rented A100 80GB or L40S works, and our guide to local versus cloud AI in India explains the trade-offs.
The other Indian models, and where they stand
Tech Mahindra Project Indus 8B
Tech Mahindra launched an 8-billion-parameter, Hindi-first education model at the summit in February 2026, built with NVIDIA under the IndiaAI Mission. It upgrades the company's earlier 1.2B Project Indus model and is aimed at teaching subjects such as physics in Hindi. Launch coverage did not say whether the weights are open or how the public can access it, so for now treat it as announced, not available.
Gnani.ai Inya VoiceOS
Gnani's Inya VoiceOS is a 5B voice-to-voice model that understands and replies in speech directly, without a separate transcription step. It covers more than 15 Indian languages and was trained on over 14 million hours of speech. It was shown as a research preview, with a 14B version planned. It is interesting for call-centre and voice-agent use, but there is no open self-serve access yet.
Soket AI Pragna 1B and Project EKA
Soket's Pragna 1B is a small 1.25B model covering English and six Indian languages, with a 2,048-token context. Soket says it releases weights under permissive licences. It is a learning and experimentation model, not a ChatGPT replacement. Soket's bigger plan, Project EKA, targets open models above 100B parameters trained on more than 2 trillion tokens of Indian data. It has not been released.
Ola Krutrim
Krutrim launched in December 2023 as India's first generative AI unicorn and released its Kruti agentic assistant in June 2025. By April 2026, MediaNama reported that Kruti had disappeared from the web and both app stores, with no official statement. The Economic Times reported that the planned Krutrim 3 model was halted, and later coverage described a pivot to AI cloud services. If you used Kruti, it is gone. It is also a useful reminder that an Indian AI app can vanish, so do not build a business on any single model without a fallback.
Bhashini
Bhashini is the government's translation and speech service for Indian languages. It is not a chat model like ChatGPT, but it is worth knowing about if your need is translation or voice in Indian languages.
What is coming next: the 12 funded projects
A MeitY reply in the Lok Sabha, reported in April 2026, listed 12 projects funded under the IndiaAI Mission's foundation models pillar. This is the pipeline to watch:
| Organisation | What it is building |
|---|---|
| BharatGen | Multilingual, multimodal models from 2B to 1T parameters for finance, legal and health |
| Sarvam AI | 120B open-source model for governance and public services |
| Soket AI | 120B open-source model for linguistic diversity, defence and healthcare |
| Fractal Analytics | 70B reasoning model for STEM and medical problem-solving |
| Gan AI | 70B multilingual model for advanced text-to-speech |
| Avataar AI | AI avatars up to 70B for agriculture and governance |
| Intellihealth | 20B model for EEG analysis and neurological screening |
| Gnani AI | 14B voice AI for real-time multilingual speech |
| Tech Mahindra Maker's Lab | 8B Indic language model and agentic platform for government |
| Shodh AI | 7B model for materials discovery |
| GenLoop | 2B small models for all 22 scheduled Indian languages |
| Zenteiq | BrahmAI, a multimodal model for engineering and scientific computing |
Notice the pattern. Only a few projects aim at a general chatbot. Most target a sector, a modality such as voice, or a specific public need. India's AI strategy is less about building one ChatGPT rival and more about models for Indian languages, voice and government services.
Which one should you use?
- You want a chat app in your language: try Indus. It is strongest for spoken Indian languages and questions about Indian life. For long documents, research or coding, ChatGPT, Gemini and Claude still do more.
- You are building an app with Indian voice or scanned documents: start with Sarvam's speech, translation and Vision APIs. The free ₹100 credit is enough to test seriously.
- You want an open model you can ship commercially: Sarvam 30B or 105B, under Apache 2.0.
- You are researching low-resource Indian languages: Param2 covers languages such as Bodo, Santali and Dogri that most models ignore. Remember the non-commercial licence.
- You are a student learning how LLMs work: small models such as Pragna 1B are good to experiment with. Our list of AI project ideas for students in India has projects you could build on them.
Language choice also affects cost. In our own measurement, the same text needed 1.65x as many tokens in Hindi as in English on the GPT-5 tokenizer, and far more on some others. Read does AI cost more in Hindi before you pick a model for Indian-language traffic.
If you want to compare Indian and global models side by side in your product, Velona gives you GPT, Claude, Gemini, DeepSeek, Qwen and many more models through one OpenAI-compatible API, paid in rupees with UPI. That makes it easy to keep a global fallback alongside an Indian model.
Frequently asked questions
Does India have its own AI model like ChatGPT?
Yes. The closest consumer equivalent is Sarvam's Indus app, which runs on the Sarvam 105B model built and trained in India. It launched in February 2026 on Android, iOS and the web. BharatGen's Param2 is also Indian-built, but it is an open model for developers, not a chat app.
Is Sarvam AI free to use?
The API gives every verified user ₹100 of free credits that do not expire. After that, sarvam-105b costs ₹15 per million input tokens and ₹60 per million output tokens, excluding GST. Sarvam did not announce a price for the Indus app at launch, so check the app listing for current terms.
Is Sarvam better than ChatGPT?
Not overall. Independent benchmarks place Sarvam 105B below leading open models on general intelligence, and it shows high hallucination rates. Sarvam is stronger in specific areas: Indian-language voice, Indian scripts and documents, and cost of Indian-language tokens.
Can I use BharatGen Param2 in a commercial app?
Not freely. Param2 is released under the BharatGen non-commercial licence, and its model card describes it as an early checkpoint that needs more fine-tuning and safety alignment. Read the licence file on Hugging Face before using it in any product.
What happened to Ola Krutrim's AI?
Krutrim's Kruti assistant went offline in April 2026 and was removed from the app stores. The Economic Times reported that the Krutrim 3 model was halted, and Krutrim has since been reported to focus on cloud services.
Which Indian AI model supports the most languages?
BharatGen Param2 lists English, Hindi and 21 more Indian languages, covering all 22 scheduled languages. Sarvam's models also target 22 Indian languages.
Where can I download Indian AI models?
Sarvam 30B and 105B are on Hugging Face under sarvamai and on AIKosh. Param2 is on Hugging Face under bharatgenai. AIKosh is the government's platform for Indian datasets and models.
Sources
- Sarvam 105B and Sarvam 30B: everything you need to know (Artificial Analysis, April 2026)
- Sarvam API pricing (Sarvam AI)
- sarvamai/sarvam-30b model card (Hugging Face)
- Sarvam AI (Wikipedia)
- India's Sarvam launches Indus AI chat app (TechCrunch, 20 February 2026)
- bharatgenai/Param2-17B-A2.4B-Thinking model card (Hugging Face)
- BharatGen launches 17B AI model in 22 Indian languages (Voice and Data, February 2026)
- BharatGen Param 2, Sarvam and the rise of Indian LLM models (Digit, February 2026)
- Tech Mahindra launches 8B Hindi-first education LLM (Analytics India Magazine, February 2026)
- Centre funds 12 AI projects for sovereign models (MediaNama, April 2026)
- Ola's Krutrim shuts down agentic AI assistant Kruti (MediaNama, April 2026)
- Are Indian-built AI apps better than ChatGPT? (eChai Ventures)