Short answer: IndiaAI foundation models are home-grown AI models built with support from the government's IndiaAI Mission. On 3 November 2026 the mission released EKA-1 and the open-source Dhīr 2B and Dhīr 5B, adding to Sarvam's 30B and 105B models. For Indian businesses they are worth piloting where Indian languages, data kept in-house or offline use matter, after testing on your own tasks.
Key takeaways
- EKA-1 from Soket has 26.3 billion parameters and covers all 22 scheduled Indian languages plus English, according to the PIB release from MeitY (2026).
- Dhīr 2B and Dhīr 5B from ZenteiQ are open-source, support English and 10 Indian languages, and are built to run on-premise or offline.
- Sarvam 30B and 105B have been open under the Apache 2.0 licence since March 2026, so they can be self-hosted.
- Published claims so far come from the government and the builders, not independent testing. Measure quality on your own data.
What did the IndiaAI Mission release in November 2026?
MeitY announced the release at a press event in New Delhi on 3 November 2026. It covers two builders selected under the mission's Innovation Centre (Foundation Models) pillar. Soket Labs built EKA-1, described as an open, sparse hybrid model trained on about 7.5 trillion tokens and meant for maths, code, reasoning and agentic use. ZenteiQ AiTech Innovations built Dhīr 2B and Dhīr 5B, compact assistants for document-grounded automation and knowledge retrieval in low-connectivity or sensitive settings, per the PIB release from MeitY (2026).
The release also says both teams used subsidised national compute from the mission, and that the models can be made available through AIKosh, the mission's catalogue of datasets and models. The mission itself was approved by the Union Cabinet in March 2024 with an outlay of ₹10,371.92 crore (about ten thousand three hundred crore rupees), according to the PIB Cabinet release (2024).
Why are IndiaAI foundation models trending now?
Three things came together. The November release added models aimed at on-device and offline use, alongside a larger reasoning model. It followed Sarvam's open release of its two models on 6 March 2026, trained in India on compute provided under the mission, as stated on the Sarvam AI blog (2026). And the question many Indian founders now ask is practical: can a model built here replace or sit beside a global API for Hindi, Marathi, Tamil or Bengali customer conversations?
Which IndiaAI-backed models can a business try today?
The table summarises what each builder states, as read in November 2026. Check each model card before you build on it, because licences and versions change.
ModelBuilderSize (as stated)LanguagesAccessLikely fit EKA-1Soket Labs26.3B parameters22 scheduled languages + EnglishDescribed as open; check the model card for licenceReasoning and agent tasks, BFSI-focused per its CEO Dhīr 2B / 5BZenteiQAbout 2B and 5BEnglish + 10 Indian languagesOpen-source, per MeitYOffline, on-premise or on-device assistants Sarvam 30BSarvam AI2.4B active parameters22 Indian languages + EnglishApache 2.0; Hugging Face and AIKoshReal-time chat and voice agents Sarvam 105BSarvam AI106B total, 10.3B active22 Indian languages + EnglishApache 2.0; Hugging Face and AIKoshComplex reasoning, coding, agent workflows Param2BharatGen (IIT Bombay-anchored)17B total, about 2.4B active22 Indian languages + EnglishOpen weights; check each model cardGeneral Indian-language text tasks The Sarvam 105B figures and its 128K-token context window come from the Hugging Face model card (2026). Param2's size is from BharatGen (2026). Sarvam says its weights are also published on AIKosh, the mission's catalogue.
What can these models do for an Indian business?
The realistic uses are the ones where language, data location or connectivity is the problem, not raw intelligence.
- Regional-language support. A WhatsApp or website assistant that answers order, appointment or policy questions in Hindi, Marathi, Tamil or Bengali, drafted from your own documents.
- Documents that should not leave your servers. Summarising contracts, claims or patient notes on hardware you control, which a self-hosted open model allows.
- Field and factory use. Offline assistants on tablets or phones for sales teams or plant staff where networks drop, which is the stated design goal of Dhīr.
- Cost control at volume. For high, steady traffic, running an open model can be cheaper than paying per call to an API, though only if you have the skills to run it.
If you already run a chatbot, our earlier guide on adding an AI chatbot to a business website covers the integration basics that apply whichever model sits behind it.
What are the honest limits?
- No independent scorecard yet. The November release describes capabilities but does not publish third-party benchmark results. Treat "advanced" claims as the builder's own.
- Coverage is not the same as quality. A model that lists 22 languages may still be much stronger in Hindi than in Odia or Assamese. Test each language your customers use.
- Self-hosting is real engineering. The Sarvam 105B model card's examples run on several GPUs at once and need extra setup for some serving tools, so it is not a laptop project.
- Open weights are not open everything. Training data and full methods are not necessarily released, and licence terms differ between models.
- You still own the outputs. Wrong answers, data handling and customer complaints remain your responsibility, so keep a human review step for anything that affects money, health or legal rights.
Should you use an Indian open model or a global AI API?
QuestionSelf-hosted Indian open modelGlobal hosted API Where does customer data go?Stays on servers you choose, including in IndiaSent to the provider; check its region and terms Indian-language focusBuilt for it, quality varies by languageOften good in Hindi, uneven in smaller languages Upfront effortHigh: GPUs or cloud setup, monitoring, updatesLow: an API key and integration Running cost patternMostly fixed infrastructurePay per use, rises with volume Offline usePossible with small models such as DhīrNot possible Many teams will end up with both: a global API for hard reasoning, and a smaller Indian model for high-volume regional-language replies or sensitive documents.
How do you pilot an IndiaAI foundation model safely?
- Pick one narrow task, such as answering delivery questions in Marathi, and write 50 to 100 real test questions with correct answers.
- Run the same set through two or three candidate models, including your current one, and score accuracy and tone in each language.
- Read the licence and the model card for each candidate, and confirm commercial use is allowed.
- Estimate hosting: try a managed GPU cloud in India before buying hardware.
- Add guardrails: retrieval from your own documents, refusal for out-of-scope questions, and logging that you review weekly.
- Launch to a small group of customers, compare complaint and handover rates, then decide.
Teams in Pune, Bengaluru or Navi Mumbai that want a second opinion can use our AI consulting and solution planning for model selection, and AI integration services to connect a chosen model to CRM, WhatsApp or support tools. Alive Inc. does not claim results for any of these models; the pilot data should decide.
Frequently Asked Questions
What is the IndiaAI Mission?
The IndiaAI Mission is the Government of India's national AI programme under MeitY, approved by the Union Cabinet in March 2024. It funds shared compute, foundation models, datasets through AIKosh, applications, skills, startup financing and safe and trusted AI, and it selected teams such as Soket and ZenteiQ to build home-grown models.
Is EKA-1 free to use for business?
MeitY describes EKA-1 as an open model, but the November 2026 release does not state its licence terms. Before any commercial use, read the licence on the official model card or the AIKosh listing, and ask your legal adviser if the terms are unclear.
Can Dhīr models run without the internet?
That is their stated design. ZenteiQ says Dhīr 2B and Dhīr 5B are lightweight enough for on-device use on mobile phones and can run without cloud connectivity, keeping data on the device. Test speed and battery use on the actual phones or tablets your staff carry.
Are Indian AI models better than ChatGPT for Hindi?
There is no neutral answer yet. Builders publish their own comparisons, and results depend heavily on the task and the language. The reliable method is a side-by-side test on your own questions, scored by fluent speakers, before you switch any customer-facing system.
Where can I download Sarvam or Param models?
Sarvam says its 30B and 105B models can be downloaded from Hugging Face and AIKosh under the Apache 2.0 licence. BharatGen lists Hugging Face and AIKosh for its Param models. Check each model card for hardware needs and licence terms.

