Indian AI models compared: Sarvam, BharatGen, Krutrim and more

AI & DeeptechIshaan VardhanNov 10, 202612 min read

If you need AI that works in Hindi, Tamil, Bengali or any of India’s 22 scheduled languages, the strongest Indian-built options in November 2026 are Sarvam’s models and APIs, BharatGen’s Param2 for research, and AI4Bharat’s free IndicTrans2 for translation. Sarvam is the only one that covers chat, speech, voice and document reading in one paid API, with open weights for its two main language models.

This guide compares what each one is, how to try it, what it costs (prices checked on the vendors’ own pages on 9 November 2026), which languages it handles and its licence, then tells you which to pick for which job, and where ChatGPT and Gemini are still ahead.

Key takeaways

• Sarvam 105B and Sarvam 30B are open-weight models under the Apache 2.0 licence, so businesses can download and run them; Sarvam’s API charges ₹15 per million input tokens and ₹60 per million output tokens for Sarvam 105B.
• For speech and documents in Indian languages, Sarvam’s Saaras v3 speech-to-text covers 23 languages and its Sarvam Vision OCR covers English plus all 22 scheduled languages.
• BharatGen’s Param2 is a 17-billion-parameter model for 22 Indian languages, but its licence is non-commercial and its makers call it an early checkpoint.
• Ola’s Krutrim-2 weights are still online, but its Kruti assistant went offline in April 2026, so it is a weak choice for new projects.
• Most benchmark wins quoted for Indian models are company-reported; an independent index still places Sarvam 105B well behind leading open models.

Indian AI models at a glance

Indian AI models differ most on licence and on what you can actually use today. The table below summarises the main options as of 9 November 2026.

Indian AI models compared. Sources: Sarvam, Hugging Face model cards, AI4Bharat, Business Today, Outlook Business; checked 9 November 2026.

Scroll the table sideways →

Sarvam: chat models, the Indus app and APIs

Sarvam is a Bengaluru AI company, founded in 2023, that builds language, speech and document models for Indian languages and sells them through an API. It had raised $41 million from Lightspeed Venture Partners, Peak XV Partners and Khosla Ventures as of February 2026, according to TechCrunch. Sarvam’s open roles are listed on our Sarvam AI jobs page.

Sarvam 105B and Sarvam 30B

Sarvam 105B and Sarvam 30B are mixture-of-experts language models, which means only part of the model runs for each word, keeping them cheaper to serve. Sarvam published both on 6 March 2026 under the Apache 2.0 licence, with weights on Hugging Face and AI Kosh, according to its launch post. The tokenizer covers all 22 scheduled Indian languages, with most training weight on the ten most spoken.

The Sarvam 105B model card lists about 106 billion parameters with 10.3 billion active per token and a 128,000-token context window. Sarvam 30B has 2.4 billion active parameters, according to Sarvam. Sarvam says its Samvaad conversational agents run on the 30B model and the Indus app on the 105B.

Both models are text-only, according to Artificial Analysis. Running Sarvam 105B yourself needs serious hardware; the model card’s examples split it across four or eight GPUs.

The Indus app

Indus is Sarvam’s consumer chat app, a ChatGPT-style assistant built on Sarvam 105B. Sarvam launched it in beta on 20 February 2026 on iOS, Android and the web, and users can type or speak and get text or audio replies, TechCrunch reported. At launch it was limited to India and could show a waitlist while Sarvam added computing capacity.

TechCrunch also noted two early limits: deleting chat history meant deleting the whole account, and the reasoning step could not be switched off, which sometimes slowed answers. Sarvam’s homepage now also promotes Indus Cowork, an agent workspace for research, data analysis, presentations and workflow automation.

Sarvam API pricing

Sarvam’s API is pay-as-you-go in rupees, and new accounts get ₹100 of free credit on signup, according to its pricing page on 9 November 2026. The main rates on that page are below.

Sarvam API pay-as-you-go prices. Source: sarvam.ai/api-pricing, checked 9 November 2026.

Scroll the table sideways →

The pricing page lists only Sarvam 105B for chat. Business-tier rates are lower for most services, and Sarvam’s documentation says its Voice Agents product starts at ₹3 a minute.

Sarvam speech, voice and document models

Sarvam’s speech and document models are the most practical Indian-built tools for work in Indian languages, because they are sold as ready APIs. According to Sarvam’s documentation on 9 November 2026:

• Saaras v3 (speech to text) handles 23 languages, the 22 scheduled Indian languages plus English, with modes for transcription, translation, transliteration and code-mixed speech such as Hinglish.
• Bulbul (text to speech): Bulbul v3 speaks 11 languages, ten Indian plus English, with adjustable pitch, pace and voice. Bulbul v3 reaches end of life on Voice Agents on 10 November 2026, and agents move to Bulbul v4.
• Mayura (translation) covers 11 languages and also transliterates; the document translation API keeps the layout of PDF and Office files.
• Sarvam Vision (document AI) reads documents in 23 languages and returns structured output.

Sarvam Vision launched on 5 February 2026 as a 3-billion-parameter vision-language model for English and all 22 scheduled languages, according to Sarvam’s launch post. Version 2.1, released on 24 September 2026, added table parsing, form extraction and handwritten Indic text, according to Sarvam.

Sarvam reports that Vision 2.1 scored 87.39 on its own Indic OCR benchmark of 6,909 samples, ahead of Gemini 3.6 Flash at 79.35, and 87.3 on the public olmOCR-Bench. These are company-reported results on a test the company built, so treat them as a starting point and test on your own documents.

BharatGen Param2: the government-backed model

Param2 is a 17-billion-parameter language model for 22 Indian languages from BharatGen, a government-backed consortium. BharatGen unveiled it at the India AI Impact Summit on 16 February 2026, and CEO Rishi Bal said roughly 4,000 GPUs were used for more than a month to train it, according to Business Today.

The Param2 model card on Hugging Face lists 17 billion parameters with 2.4 billion active, a 32,768-token context window, and English, Hindi and 21 other Indian languages. It was trained on about 22.5 trillion tokens. The card says a 128,000-token version is coming.

Two limits matter. The licence is the BharatGen non-commercial licence, so check its terms before using Param2 in a product. And the card calls this release an early post-training checkpoint that needs more fine-tuning and safety work before production use. No hosted app or API is listed; you run it yourself.

The card’s own benchmark table is modest: an MMLU score of 57.79 sits below the comparison models it lists. Bal told Business Today that most business and government uses do not need a frontier-sized model, and that Param2 will be the base for later voice and vision systems.

Krutrim: still downloadable, no longer a consumer app

Krutrim is Ola founder Bhavish Aggarwal’s AI company, and its Krutrim-2 model remains available but its consumer assistant does not. The Krutrim-2 model card describes a 12-billion-parameter model built on Mistral’s NeMo architecture with a 128,000-token context window, released under the Krutrim Community License.

Kruti, the assistant built on that model, went offline on app stores, the web and inside the Ola app in April 2026 with no stated reason, Outlook Business reported on 17 April 2026, after layoffs that included regional-language transcribers. Krutrim’s website now leads with cloud computing, including code sandboxes for AI agents from ₹1.56 an hour, rather than with its models.

Our view: Krutrim-2 is worth a look only if you are already on Krutrim’s cloud. For new Indian-language projects, Sarvam’s open models have a clearer licence and active development.

AI4Bharat and Gnani: translation and voice

AI4Bharat’s IndicTrans2 is the best free option for translation between English and Indian languages. According to its GitHub repository, IndicTrans2 is an open-source translation model for all 22 scheduled languages, translating English to Indic, Indic to English and between Indian languages. Both the code and the model checkpoints are under the MIT licence, which allows commercial use. The main models have 1 billion parameters, with smaller distilled versions of 200 million and 320 million. You can try it on AI4Bharat’s demo site, on Hugging Face or in a Colab notebook.

Gnani.ai is a Bengaluru voice AI company that sells mainly to enterprises. It unveiled Inya VoiceOS, a 5-billion-parameter model that works on speech directly rather than converting it to text first, at the February 2026 summit, released as a research preview with support for more than 15 Indian languages, according to Business Today. Gnani’s website now lists its Prisma speech-to-text model and claims coverage of 40-plus languages and dialects. It offers an API key signup and demos, but no public prices.

Which Indian AI to use for what

The right Indian AI depends on the job: Sarvam for most paid API work, IndicTrans2 for free translation, and Param2 only for research. Here is our pick for each common task, based on the facts above.

Which Indian AI to use for what. 100xFounder’s picks, based on vendor pages checked 9 November 2026.

Scroll the table sideways →

Where ChatGPT and Gemini still lead

ChatGPT and Gemini still lead on general reasoning, coding, multimodal features and reach, even though Indian models win some Indian-language tests. In January 2026, ChatGPT had 180 million monthly active users in India and Gemini 118 million, according to Sensor Tower data reported by TechCrunch; no Indian app has published comparable figures.

On benchmarks, the gap is clearer in independent tests than in company ones. Sarvam’s own tables show Sarvam 105B beating several open models, but they also show OpenAI’s o4-mini ahead on the AIME 25 maths test and LiveCodeBench coding test. Artificial Analysis, which runs its own tests, gave Sarvam 105B a score of 18 on its Intelligence Index on 2 April 2026, against 33 for OpenAI’s open gpt-oss-120B, and flagged high hallucination rates for both Sarvam models.

Sarvam also says Sarvam 105B won 90% of pairwise comparisons on Indian-language prompts judged by Gemini 3, without naming the models it was compared against. Claims like that, and Param2’s own benchmark table, are useful signals but not proof. For anything important, run your own test set in the languages your users speak.

Bottom line

For most founders and teams in November 2026, Sarvam is the practical Indian choice: open weights under a business-friendly licence, a single API for text, speech and documents, and rupee pricing. IndicTrans2 is the free pick for translation, Param2 is one to watch for research, and Krutrim is hard to recommend for new work. Use ChatGPT or Gemini where you need the strongest general reasoning, and an Indian model where language coverage, cost or keeping data in India matters most. If you want to build these systems, browse current AI and ML jobs, and for the market backdrop read how India became ChatGPT’s largest market.

FAQ

Which is the best Indian AI model right now?

For most uses, Sarvam 105B, because it is open-weight under Apache 2.0, powers the Indus app and is available through a paid API with rupee pricing. It covers the 22 scheduled Indian languages. Independent tests still rank it below leading open models from abroad, so compare it with ChatGPT or Gemini on your own tasks.

Is Sarvam AI free to use?

The Indus app is free to try in beta, and new API accounts get ₹100 of free credit, according to Sarvam’s pricing page on 9 November 2026. After that, the API is paid: Sarvam 105B costs ₹15 per million input tokens and ₹60 per million output tokens. The model weights are free to download under Apache 2.0.

Which AI is best for speech to text in Indian languages?

Sarvam’s Saaras v3 is the easiest Indian option: it handles 23 languages, including code-mixed speech like Hinglish, and costs ₹0.49 a minute on pay-as-you-go, according to Sarvam’s documentation and pricing page on 9 November 2026. Large call-centre users can also talk to Gnani.ai, which does not publish prices.

Can I use BharatGen Param2 in a commercial product?

Not without checking carefully. Param2’s Hugging Face model card lists a BharatGen non-commercial licence, and its makers describe the release as an early checkpoint that needs more fine-tuning and safety alignment before production. For commercial use, Sarvam’s Apache 2.0 models or the MIT-licensed IndicTrans2 are clearer choices.

What happened to Ola’s Krutrim AI app?

Kruti, Krutrim’s AI assistant, went offline on app stores, the web and inside the Ola app in April 2026 without an explanation, Outlook Business reported on 17 April 2026. The Krutrim-2 model weights remain on Hugging Face, and Krutrim’s website now focuses on cloud computing services.

Sources

1. Sarvam, “Sarvam 30B and 105B”, 6 March 2026. Link
2. Hugging Face, “sarvamai/sarvam-105b” model card, checked 9 November 2026. Link
3. Sarvam, “API pricing”, checked 9 November 2026. Link
4. Sarvam, “Documentation”, checked 9 November 2026. Link
5. Sarvam, homepage, checked 9 November 2026. Link
6. Sarvam, “Sarvam Vision”, 5 February 2026. Link
7. Sarvam, “Sarvam Vision 2.1”, 24 September 2026. Link
8. TechCrunch, “India’s Sarvam launches Indus AI chat app as competition heats up”, 20 February 2026. Link
9. Artificial Analysis, “Sarvam 105B and Sarvam 30B: everything you need to know”, 2 April 2026. Link
10. Hugging Face, “bharatgenai/Param2-17B-A2.4B-Thinking” model card, checked 9 November 2026. Link
11. Business Today, “‘Most real-world use cases don’t need trillion-parameter AI’: BharatGen CEO Rishi Bal on Param 2 launch”, 16 February 2026. Link
12. Hugging Face, “krutrim-ai-labs/Krutrim-2-instruct” model card, checked 9 November 2026. Link
13. Outlook Business, “Bhavish Aggarwal’s AI bet hits another snag as Kruti goes offline”, 17 April 2026. Link
14. Krutrim, homepage, checked 9 November 2026. Link
15. AI4Bharat, “IndicTrans2” GitHub repository, checked 9 November 2026. Link
16. Business Today, “India AI Impact Summit 2026: Gnani.ai launches 5-billion-parameter voice AI model under IndiaAI Mission”, 17 February 2026. Link
17. Gnani.ai, homepage, checked 9 November 2026. Link
18. TechCrunch, “India’s AI boom pushes firms to trade near-term revenue for users”, 24 February 2026. Link

Related stories

AI & DeeptechNvidia drops below $4 trillion as Congress presses Jensen HuangIshaan VardhanNov 14, 20266 min read

AI & DeeptechIndiaAI Mission 2026: GPU compute, foundation models and how to applyIshaan VardhanOct 18, 202611 min read

Sources (18)

1. Sarvam 6 Mar 202630B/105B: MoE, 2.4B active (30B), Apache 2.0, HF + AI Kosh, 22 languages tokenizer, Samvaad/Indus, company benchmarks incl. o4-mini lead, 90% pairwise win
2. Hugging Face (Sarvam) 9 Nov 2026105B: ~106B total, 10.3B active, 128K context, Apache 2.0, TP examples
3. Sarvam 9 Nov 2026API prices in INR, ₹100 signup credit
4. Sarvam 9 Nov 2026Saaras v3 23 langs; Bulbul v3 11 langs, EOL 10 Nov on Voice Agents; Mayura 11; Vision 23; Voice Agents from ₹3/min
5. Sarvam 9 Nov 2026Indus Cowork agent workspace
6. Sarvam 5 Feb 2026Sarvam Vision 3B, 22 languages + English
7. Sarvam 24 Sep 2026Vision 2.1: tables, forms, handwriting; Indic OCR bench 87.39 vs Gemini 3.6 Flash 79.35; olmOCR 87.3; 6,909 samples
8. TechCrunch 20 Feb 2026Indus beta launch, platforms, waitlist, limits; Sarvam founded 2023, $41M raised from Lightspeed, Peak XV, Khosla
9. Artificial Analysis 2 Apr 2026Intelligence Index 105B=18, gpt-oss-120B=33; text-only; hallucination rates
10. Hugging Face (BharatGen) 9 Nov 2026Param2 17B/2.4B active, 32K context, 23 languages, ~22.5T tokens, non-commercial licence, early checkpoint, MMLU 57.79
11. Business Today 16 Feb 2026Param2 unveiled 16 Feb 2026; ~4,000 GPUs; Rishi Bal comments
12. Hugging Face (Krutrim) 9 Nov 2026Krutrim-2 12B Mistral-NeMo, 128K, Krutrim Community License
13. Outlook Business 17 Apr 2026Kruti offline across platforms; layoffs
14. Krutrim 9 Nov 2026Cloud focus; sandbox from ₹1.56/hr
15. AI4Bharat 9 Nov 2026IndicTrans2: 22 languages, MIT licence, 1B/200M/320M, demo/HF/Colab
16. Business Today 17 Feb 2026Gnani Inya VoiceOS 5B, 15+ Indian languages, research preview
17. Gnani.ai 9 Nov 2026Prisma STT; 40+ languages and dialects claim; API signup; no public pricing
18. TechCrunch (Sensor Tower data) 24 Feb 2026Jan 2026 MAU ChatGPT 180M, Gemini 118M in India