EKA-1 with 26.3 Bn parameters leads India’s new homegrown AI models under IndiaAI Mission

Two Indian startups have unveiled foundation models under the IndiaAI Mission. Soket Labs showed EKA-1, and ZenteiQ showed Dhīr 2B and Dhīr 5B. Inc42, Mint and The Economic Times reported the launch. S Krishnan, MeitY Secretary, was present, according to Inc42 and The Economic Times.

EKA-1 has 26.3 Bn parameters and was trained on nearly 7.5 Tn tokens, first reported by Inc42. It is an open, sparse hybrid model. It supports all 22 scheduled Indian languages and English, and targets maths, code, reasoning and agentic uses. Abhishek Upperwal, cofounder and CEO of Soket Labs, said, as first reported by Inc42: "With the support of the IndiaAI Mission, we have developed an advanced AI model specifically tailored to India’s unique banking, financial services and cybersecurity landscape."

The Dhīr models are smaller, at two billion and five billion parameters. They cover English and 10 other Indian languages. Professor Sashikumaar Ganesan, CEO of ZenteiQ, said, according to The Economic Times: "We have launched two small models. We call them compact models". Such models suit offline, on-device use where connectivity is limited. Both startups were picked under the IndiaAI Innovation Centre pillar, and subsidised compute helped their work.

Krishnan said indigenous models will add to India's technological and strategic autonomy and lift productivity, as first reported by Inc42. Inc42 also noted that no benchmark results, deployment details or pricing were provided.