On Nov. 3, 2026, the IndiaAI Mission under the Ministry of Electronics and Information Technology unveiled two indigenous foundation-model families at a press event in New Delhi attended by MeitY Secretary S. Krishnan: EKA-1 from Soket Labs Technology and Research, and open-source Dhīr 2B and Dhīr 5B from ZenteiQ AiTech Innovations.
Key facts
| Item | Detail |
|---|---|
| Released | Nov. 3, 2026, New Delhi |
| Program | IndiaAI Mission, MeitY |
| EKA-1 | 26.3B parameters, ~7.5T tokens, 22 scheduled Indian languages plus English |
| Dhīr 2B | 2.03B parameters, dense |
| Dhīr 5B MoE | 5.09B total, 1.79B active |
| Dhīr languages | English plus 10 Indian languages |
| Dhīr licence | Apache 2.0 |
What was released
The Press Information Bureau described the event as the IndiaAI Mission releasing indigenous foundation models under the IndiaAI Innovation Centre (Foundation Models) pillar. Both teams were selected under that pillar. EKA-1 targets broad multilingual capability across India's scheduled languages; the Dhīr pair targets compact, deployable assistants for sensitive or disconnected settings.
EKA-1 by Soket
PIB calls EKA-1 “a powerful Foundation Model featuring 26.3 billion parameters and trained on approximately 7.5 trillion tokens,” spanning all 22 scheduled Indian languages alongside English. It is “Built as an open, sparse hybrid language model, EKA-1 is designed for math, code, reasoning, and agentic use.”
Soket Labs CEO Abhishek Upperwal said in the release: “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.”
Dhīr 2B and 5B by ZenteiQ
PIB says Dhīr 2B and Dhīr 5B are open-source models that “can be deployed as secure, on-premise, multilingual assistants supporting English and 10 Indian languages,” enabling “offline-capable, document-grounded automation, knowledge retrieval, and citizen or enterprise services across defence, government, industrial, and educational sectors particularly in low-connectivity or sensitive environments.”
ZenteiQ describes them on its product page as “Compact foundation models pretrained from scratch in India, for offline use on laptops, embedded boards and edge devices, in English and ten Indian languages.” Both ship under Apache 2.0 with downloads on AIKosh. Dhīr 2B is a dense transformer with 2.03B parameters, 16,384-token context, and a 4-bit file of 1.21 GB. Dhīr 5B MoE activates 2 of 12 experts per token, with 5.09B total and 1.79B active parameters per token, 16,384 context, and a 4-bit file of 2.95 GB. Pretraining used “6T+” tokens per model. Languages include Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi and Odia, alongside English. Every training stage ran on Google Cloud TPU v6e in Mumbai.
Company-reported overall scores are 61.4 for Dhīr 2B and 64.6 for Dhīr 5B MoE; ZenteiQ says Dhīr 2B is first in its class on 13 of 15 Indic benchmarks (company-reported). For context on small open models built for devices, see Liquid AI's LFM3 Puts Six Open 2B-Class Models on Devices and Moonshot's Kimi K3.1 Pro Is a 3T-Class Open-Weight Model.
Running offline on small hardware
Prof. Sashikumar, CEO of ZenteiQ, said in PIB: “Their lightweight architecture also enables on-device deployment on mobile phones, without dependence on cloud connectivity.” ZenteiQ lists decode speeds in tokens per second at 4-bit with llama.cpp (company-reported): on a Raspberry Pi 4 CPU, 2.3 for Dhīr 2B and 2.8 for Dhīr 5B; on NVIDIA Jetson Orin NX, 34.1 and 39.9; on Intel Iris Xe laptop graphics, 18.6 and 20.5.
Limits ZenteiQ lists
ZenteiQ states verbatim: “Neither model has a reasoning mode, and we make no claim of scientific or engineering reasoning ability.” It also notes: “Text only: no image, audio or speech input or output.” Dhīr 9B is “In development; public release subject to evaluation.”
How the IndiaAI Mission helped
PIB says access to “subsidized national AI compute infrastructure and support for ancillary research expenses under the mission enabled the development of these models.” Both Soket and ZenteiQ were chosen under the Innovation Centre foundation-model track.
What it means
Analysis (ours). The Nov. 3 release pairs two different bets: EKA-1's scale and 22-language span for high-end reasoning and agentic workloads tied to regulated sectors Soket names, against ZenteiQ's Apache-licensed edge stack aimed at air-gapped or low-bandwidth deployment. Labelled company benchmarks and llama.cpp speed tables help buyers compare Dhīr to other small open models, but ZenteiQ's explicit caps on reasoning and multimodal I/O set expectations for where these weights fit in a stack versus cloud foundation models.
What to watch
Whether EKA-1 moves from announcement to accessible weights or APIs, how AIKosh downloads perform in government pilots, and whether Dhīr 9B clears ZenteiQ's evaluation gate for a public release.