Three Indian foundation models unveiled in New Delhi on November 3 are designed to keep working where cloud access is limited or sensitive data must stay on local equipment. Two of them, Dhīr 2B and Dhīr 5B, can run on a smartphone with a minimum of 6GB RAM after installation.
ZenteiQ AiTech Innovations developed the Dhīr models. Soket Labs Technology and Research built the third, EKA-1, for Indian languages, mathematics, coding, reasoning and agentic applications.
Different models for different workloads
ZenteiQ’s open-source Dhīr 2B and Dhīr 5B assistants offer multilingual support spanning 10 Indian languages and English. ZenteiQ says they can perform document-grounded automation and knowledge retrieval for citizen and enterprise services, including defence, government, industrial and educational work.
The models can operate on GPUs and TPUs as well as mobile devices. Their local operation means queries and documents can remain on the user’s phone or an organisation’s premises rather than travel to an external cloud service.
“Once downloaded, all the configuration automatically takes place, then the model will be sitting in the user’s mobile itself,” ZenteiQ founder and chief scientist Sashikumaar Ganesan told ANI. The company is developing a free Android and iOS application, while Dhīr’s model weights are already available through AIKosh.
EKA-1 is substantially larger, with a parameter count of 26.3 billion; its training corpus contained approximately 7.5 trillion tokens. Soket describes an open, sparse hybrid model covering all 22 scheduled Indian languages alongside English.
Its stated targets include banking, financial services and cybersecurity. The model can be deployed on private servers, private clouds and environments without internet access, although Soket says only certain inferencing can take place locally on a smartphone.
Public compute behind private development
The IndiaAI Innovation Centre chose both companies for its Foundation Models pillar. The programme supplied subsidised national computing capacity and assistance with additional research costs.
Government disclosures value Soket’s support at ₹177.08 crore: ₹162.47 crore for compute and ₹14.61 crore for other research. ZenteiQ received ₹206.49 crore, with non-compute assistance accounting for ₹41.30 crore and computing support for ₹165.19 crore.
MeitY Secretary S. Krishnan attended the launch with ministry and IndiaAI officials. He said indigenous models could strengthen technological and strategic autonomy while supporting productivity across government, industry, research and education.
Soket CEO Abhishek Upperwal linked EKA-1 to India’s particular financial and security requirements. “We want to remove the reliance of foreign models and we want to make an indigenous model which can work in all our critical sectors,” he said.
Company co-founder and chief science officer Mayank Singh said its data preparation, training and evaluation were carried out entirely in India. Soket plans to release EKA-1’s weights openly within a month, initially through a waitlist.
Access and what comes next
AIKosh is intended to put the models within reach of researchers, students, developers and businesses. EKA-1 and the Dhīr systems join earlier IndiaAI-supported releases from Sarvam AI, BharatGen and Gnani AI.
ZenteiQ also plans a December release aimed at researchers and engineers addressing specialised problems, using a model with 9 billion parameters. IndiaAI chief operating officer Sudeep Shrivastava described the current releases as spanning lightweight systems for everyday devices and larger models for complex enterprise applications.
The commercial backdrop is expanding as well. NASSCOM and Boston Consulting Group estimate that India’s AI services market could reach $17 billion by 2027.
Sources
- IndiaAI Mission: Indian startups launch offline AI models Dhīr and EKA-1
- Artificial intelligence industry in India
- IndiaAI Mission Launches 3 Indigenous AI Foundation Models
- IndiaAI Mission Releases 3 Indigenous AI Models
- IndiaAI Mission's three indigenous AI models target offline mobile access, secure use in critical sectors
- IndiaAI Mission Boosts National Autonomy With New Indigenous AI Models
Photo: Scott Rodgerson / Unsplash (Unsplash License)
