The IndiaAI Mission has released three home-grown foundation models. EKA-1, from Bengaluru’s Soket Labs, is a 26-billion-parameter multilingual model. Dhīr 2B and Dhīr 5B, from ZenteiQ AiTech, are compact open-weight models designed to run without the internet. The releases show that India’s state-backed model programme is producing results. How good the models are is still unknown, because no benchmark results have been published.
When the Union Cabinet approved the IndiaAI Mission in March 2024, with an outlay of ₹10,371.92 crore, its most ambitious promise was that India would build its own foundation models: large AI systems trained on Indian data and languages, owned and governed in India. On Tuesday, two more companies backed by the mission showed what they had built.
The Ministry of Electronics and Information Technology (MeitY) released the models in New Delhi, PTI reported. MeitY Secretary S. Krishnan said indigenous models strengthen the country’s “technological and strategic autonomy”. NSH did not find a PIB or MeitY press release, so the details below rely on PTI and other news reports and on the companies’ own descriptions.
What was released
EKA-1, from Soket Labs (also called Soket AI Labs), is the larger of the two families. According to the reports:
- It has 26.3 billion parameters, the adjustable numerical weights a model learns during training
- It was trained on about 7.5 trillion tokens, the word fragments that models read
- It is described as an “open, sparse hybrid” model for mathematics, code, reasoning and “agentic” tasks, in which an AI carries out multi-step actions
- It covers all 22 languages in the Eighth Schedule of the Constitution plus English, according to most reports. The Tribune instead described coverage of “more than 100 Indian and international languages”
A sparse model activates only part of its network for each word it processes. That cuts computing cost and is one reason Soket says some inference can run on a device. Chief executive Abhishek Upperwal named banking, financial services and insurance (BFSI) and cybersecurity as target sectors. The licence for EKA-1 was not stated in the material NSH reviewed.
Dhīr 2B and Dhīr 5B, from ZenteiQ AiTech Innovations, take a different approach. They are small models, sized by their names at about 2 and 5 billion parameters, though exact counts were not published. The Tribune described them as mixture-of-experts designs, in which different sub-networks specialise in different inputs. They support English and ten Indian languages and are built for offline, on-premise and even air-gapped deployment, meaning computers kept physically disconnected from outside networks. ZenteiQ says the weights are released under the Apache 2.0 licence, which allows commercial use and modification. The models are being distributed through AIKosh, the mission’s platform for datasets and models; NSH did not verify the listings themselves.
Why small and offline matters
The global race in AI has been dominated by ever-larger models served from vast data centres. For India, that model has drawbacks. Connectivity is uneven outside the cities. Government departments, hospitals, banks and defence users often cannot send sensitive data to foreign cloud servers. Many users work in languages that global models handle less well.
There is also a technical reason Indian languages need dedicated models. Models read text as “tokens”, and tokenisers designed mainly for English tend to break words in Indic scripts into many more pieces. That makes every sentence in Telugu or Hindi more expensive to process and leaves less room in the model’s working memory. Models trained with Indian text from the start can use tokenisers that handle these scripts more efficiently.
A model that runs on a laptop, a hospital server or even a phone, in Telugu or Odia, addresses all three problems at once. It is also cheaper to adapt: a 2- or 5-billion-parameter model can be fine-tuned on a modest budget by a university laboratory or a state agency. For a Telangana district office or an Andhra Pradesh agriculture extension service, an open model that can be installed locally may be more useful than a more capable one that cannot.
Where it fits
The mission’s foundation-model pillar has selected twelve organisations, according to a PIB release of 13 February 2026. Besides Soket and ZenteiQ, they are Sarvam, Gnani, Gan, Avatar, the IIT Bombay-led BharatGen consortium, GenLoop, Intellihealth, Shodh AI, Fractal and Tech Mahindra Makers Lab. The same release said more than 38,000 graphics processors had been made available to developers at subsidised rates. Sarvam, chosen first in April 2025, has already released models; Tuesday’s releases broaden the field.
For readers in Hyderabad, the relevance is direct. The city has one of India’s largest concentrations of AI engineering talent, a growing data-centre cluster and state-government ambitions in AI. Open-weight Indian models give its start-ups and institutions a base to build on without depending entirely on foreign providers.
The caveats
Three questions remain open.
- Performance. No benchmark results were reported. Until independent evaluations appear, including tests in Indian languages, claims about reasoning or coding ability are the developers’ own. Size alone says little: a well-trained smaller model can outperform a poorly trained larger one.
- Openness. “Open” is used loosely across the industry. Dhīr’s Apache 2.0 licence is clear; EKA-1’s terms were not stated, and the training data have not been described in detail.
- Safety and governance. One report mentioned safeguards on self-harm content, but NSH could not confirm it independently. How these models handle misinformation, bias and misuse in Indian contexts will matter as much as their scores.
The comparison with frontier systems is not the right test. India’s mission does not need to beat the world’s largest models. It needs to produce dependable, affordable systems that Indian institutions can inspect, adapt and run on their own terms.
