Soket Labs and ZenteiQ AiTech Innovations debuted three new large language models under the IndiaAI Mission. Soket's EKA-1 spans 22 Indian languages with 26.3 billion parameters, while ZenteiQ's Dhīr models focus on smaller, on-device multilingual assistants. This pushes India's ambition for technological self-reliance in AI, aiming to move beyond reliance on global foundational models.
How We Got Here
The IndiaAI Mission selected Soket Labs and ZenteiQ to develop domestic foundation models, providing subsidised AI compute and research support. These models were unveiled yesterday in New Delhi, with IT secretary S Krishnan present.
The Numbers
- Soket's EKA-1 is a 26.3 billion parameter model, trained on 7.5 trillion tokens, supporting all 22 scheduled Indian languages and English.
- EKA-1 targets mathematics, code generation, reasoning, and agentic applications, with specific focus on banking, financial services, and cybersecurity.
- ZenteiQ debuted Dhīr 2B and Dhīr 5B, which are open-source multilingual models supporting English and 10 other Indian languages.
- The Dhīr models offer on-premise and on-device deployment on mobile phones, allowing local AI capabilities without cloud connectivity for enhanced privacy.
- The announcement provided no benchmark results, deployment specifics, or pricing information for any of the new models.
What Happens Next
🇮🇳 Why This Matters for India
For SaaS founders in Ahmedabad building for local governance or fintech in Chennai, access to privacy-focused, offline-capable Indian language models could significantly lower infra costs and data privacy hurdles.
The Take
The actual utility of these models for Indian startups hinges entirely on performance benchmarks and competitive pricing versus established open-source models like Llama. Until those numbers are out, this feels more like a strategic statement about domestic capability than a clear path to widespread developer adoption.
Source:
Inc42 ↗