
Tech & AI
By TrendScope Daily Team · November 16, 2026
India now has a homegrown large language model built from scratch for its own languages. EKA-1, a 26.3-billion-parameter AI model from Bengaluru-based Soket Labs, was unveiled by the Ministry of Electronics and Information Technology (MeitY) on November 3, 2026, alongside two smaller sibling models called Dhīr. Together they are India’s clearest answer yet to the question of whether the country can build serious AI instead of only using AI built elsewhere.
What exactly is EKA-1?
EKA-1 is described by its maker, Soket Labs, as an “open, sparse hybrid language model.” It was trained on roughly 7.5 trillion tokens and supports English plus all 22 languages scheduled in the Indian Constitution — from Hindi and Tamil to Assamese and Dogri. Soket Labs built it under the IndiaAI Mission’s Innovation Centre (Foundation Models) pillar, which gives selected startups subsidised national AI compute and research support so they can train models without needing the billion-dollar budgets of Silicon Valley labs.
Soket Labs co-founder and CEO Abhishek Upperwal has said the model was built with India’s banking, financial services and cybersecurity sectors specifically in mind — the kind of regulated, high-stakes use cases where a company may not want its data processed on a foreign server.
Meet the smaller siblings: Dhīr 2B and Dhīr 5B
Alongside EKA-1, Hyderabad-based ZenteiQ released two lighter models, Dhīr 2B and Dhīr 5B, named after the Sanskrit word for steady or resolute. Both are mixture-of-experts models trained from scratch on Google Cloud TPUs, and both support English plus 10 Indian languages. ZenteiQ says the real selling point is that they can run offline — on a GPU, a TPU, or even a mobile phone with no internet connection — which matters for government, defence and rural deployments where connectivity isn’t guaranteed.
| Model | Developer | Size | Languages | Runs Offline | Built For |
|---|---|---|---|---|---|
| EKA-1 | Soket Labs | 26.3B parameters | English + 22 scheduled languages | Not stated | Banking, finance, cybersecurity, reasoning |
| Dhīr 2B | ZenteiQ | 2B parameters | English + 10 Indian languages | Yes, incl. on mobile | Lightweight on-device assistant |
| Dhīr 5B | ZenteiQ | 5B parameters | English + 10 Indian languages | Yes, incl. air-gapped setups | Secure enterprise & research use |
“This initiative will strengthen India’s strategic autonomy in artificial intelligence.”
— S Krishnan, Secretary, Ministry of Electronics and Information Technology
Why it matters
India has talked for years about building “sovereign AI” — models that are trained on Indian data, tuned for Indian languages, and not dependent on a handful of American or Chinese labs. EKA-1 and the Dhīr models are the IndiaAI Mission’s most concrete proof of that so far, following the big foreign AI platforms already racing for a share of the Indian market. The government says the models will be made available to researchers, students and developers through its AIKosh platform, with ZenteiQ releasing the Dhīr model weights under the permissive Apache 2.0 licence.
It also lands just as the leadership of the global AI industry is in flux — coming only days after Nvidia’s own leadership shake-up renewed questions about who controls the AI supply chain that India’s own models will eventually need to run on.
What to watch next
- Independent benchmark scores comparing EKA-1 and Dhīr against GPT, Gemini and Llama-class models.
- Whether banks and fintechs actually adopt EKA-1 for regulated workloads, as Soket Labs intends.
- How quickly developers outside government and research institutions can access the models on AIKosh.
If you want to go deeper into how large language models like these are trained and evaluated, a plain-language primer is useful background — for instance Kai-Fu Lee and Chen Qiufan’s AI 2041, which lays out in accessible terms how national AI strategies like India’s are likely to play out.
As an Amazon Associate, TrendScope Daily earns from qualifying purchases made through links in this article.
Quick questions
What is EKA-1 in simple terms?
EKA-1 is a large AI language model built in India by Soket Labs, with 26.3 billion parameters, that understands English and all 22 constitutionally scheduled Indian languages.
Can the public use EKA-1 or Dhīr right now?
The government says the models will be made available through its AIKosh platform for researchers, students and developers; Dhīr’s weights are released under the Apache 2.0 open-source licence. Check AIKosh for current access terms before use.
How is EKA-1 different from ChatGPT or Gemini?
EKA-1 was trained specifically for Indian languages and use cases like banking and cybersecurity, and built with subsidised Indian government compute support, unlike foreign models trained primarily on global, largely English-centric data.
Related on TrendScope Daily
- What Is Indus AI? Sarvam’s India-Made ChatGPT Rival, Explained
- Flipkart Inside Gemini: How Google’s AI Shopping Test Works
- Why Did Jensen Huang Resign as Nvidia CEO? India’s AI Stakes
- More Tech & AI stories: browse the Tech & AI section
- All our AI explainers: AI Explained hub
- Today’s full roundup: Trending Today in India: November 16, 2026
Sources: IndiaAI Mission press coverage, Swarajya.