The IndiaAI Mission, run under MeitY, released three indigenous foundation models on November 3, 2026, at an event in New Delhi attended by MeitY Secretary S. Krishnan, according to the Press Information Bureau (PIB). The largest is EKA-1 from Soket Labs, a 26.3 billion parameter model trained on roughly 7.5 trillion tokens. The other two, Dhir 2B and Dhir 5B from ZenteiQ, are small enough to run offline on modest hardware. For anyone choosing among these IndiaAI foundation models: EKA-1 is the one to watch for banking and security work once its weights open, while Dhir is the one you can deploy today.
TL;DR
- Three models were released on November 3, 2026: EKA-1 (Soket Labs), Dhir 2B and Dhir 5B (ZenteiQ), per PIB.
- EKA-1 has 26.3B parameters, about 7.5T training tokens, and covers all 22 scheduled Indian languages plus English.
- Dhir 2B and 5B are Apache 2.0, distributed through AIKosh, in English plus 10 Indian languages.
- EKA-1 weights are not yet public; an open release is planned within a month, waitlist first.
- Dhir's llama.cpp speeds make laptop and Jetson hardware practical; a Raspberry Pi 4 is a demo only.
What exactly did MeitY release under the IndiaAI Mission?
Three models from two companies. PIB describes EKA-1 as an open, sparse hybrid language model built by Soket Labs Technology and Research Private Limited for maths, code, reasoning and agentic use. Dhir 2B and Dhir 5B come from ZenteiQ AiTech Innovations and are positioned as secure, on-premise, offline-capable multilingual assistants for defence, government, industry and education, particularly where connectivity is poor or data is sensitive.
Coverage in The Hindu and Moneycontrol matches the PIB account. Soket CEO Abhishek Upperwal said EKA-1 was tailored to India's banking, financial services and cybersecurity landscape.
How do EKA-1, Dhir 2B and Dhir 5B compare on specifications?
| Model | Parameters | Architecture | Context | 4-bit file | Languages | Licence status |
|---|---|---|---|---|---|---|
| EKA-1 | 26.3B (PIB) | Open, sparse hybrid | Not stated in the release | Not published | 22 scheduled languages plus English | Weights planned within a month, waitlist first |
| Dhir 2B | 2.03B (ZenteiQ) | Dense transformer | 16,384 tokens | 1.21 GB | English plus 10 Indian languages | Apache 2.0 |
| Dhir 5B | 5.09B total, 1.79B active per token (ZenteiQ) | Mixture of experts, 2 of 12 experts active | 16,384 tokens | 2.95 GB | English plus 10 Indian languages | Apache 2.0 |
The Dhir languages are Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi and Odia. The 16-bit files are larger, at 4.06 GB for Dhir 2B and 10.19 GB for Dhir 5B, which is why the 4-bit builds matter for edge devices.
Is EKA-1 open source yet?
Not yet, and the distinction matters. PIB uses the word "open" in the architecture description, but the weights themselves are not downloadable today. Soket chief science officer Mayank Singh said an open release is planned within a month, with initial access through a waitlist, as reported by India Gazette. Until then, treat EKA-1 as announced rather than available.
Can Dhir really run offline, and on what hardware?
Yes. ZenteiQ publishes single-stream decode speeds using llama.cpp with 4-bit weights (ZenteiQ):
| Hardware | Dhir 2B | Dhir 5B MoE |
|---|---|---|
| Intel Iris Xe laptop graphics | 18.6 tok/s | 20.5 tok/s |
| NVIDIA Jetson Orin NX | 34.1 tok/s | 39.9 tok/s |
| Raspberry Pi 4 (CPU) | 2.3 tok/s | 2.8 tok/s |
At 18 to 20 tokens per second on laptop graphics, interactive chat is comfortable; a Jetson Orin NX leaves headroom for kiosks or field devices; a Raspberry Pi 4 at 2.3 tok/s suits a proof of concept over a working tool. All are vendor-measured, single-user figures, so test with your own prompts.
The 5B is faster than the 2B in every row because its mixture-of-experts design activates only 1.79B parameters per token (ZenteiQ), though it still holds all 5.09B in memory.
How good are the Dhir models, and who tested them?
ZenteiQ evaluated both models against 24 open models in the 1 to 7B range across 41 benchmarks under a single protocol (ZenteiQ). Overall scores were 61.4 for Dhir 2B and 64.6 for Dhir 5B. Section scores, in percent:
| Section | Dhir 2B | Dhir 5B |
|---|---|---|
| Indic | 53.7 | 51.2 |
| Agentic | 85.9 | 87.2 |
| Code and maths | 50.8 | 61.2 |
| English | 55.1 | 58.9 |
ZenteiQ also reports that Dhir 2B ranks first in its class on 13 of 15 Indic benchmarks, so the smaller model beats the larger on Indian-language work. Pick the 2B for Indian-language text and the 5B for code and maths. These are developer-side results; await third-party replication before treating them as settled.
Where were the models trained, and does that matter?
Both Dhir models were pretrained from scratch, from random initialisation rather than adapted from another developer's checkpoint, on Google Cloud TPU v6e in Mumbai, using 2,048 TPU v6e chips and over 6 trillion tokens each, with an in-house 131k-entry tokenizer (ZenteiQ). ZenteiQ lists 64 NVIDIA H100 GPUs as in production (serving, not training). The company was founded out of IISc Bangalore (ZenteiQ).
For regulated buyers the point is data locality: a model that fits on a laptop or edge box removes per-call cloud inference charges and keeps prompts inside your network, unlike a subscription assistant such as Claude Pro pricing, where every query leaves the building.
How do these IndiaAI foundation models fit the wider mission?
The IndiaAI Mission was approved by the Union Cabinet in March 2024 with an outlay of Rs 10,372 crore (PIB). It works across seven pillars: compute capacity, an innovation centre, the AIKosh datasets platform, application development, FutureSkills, startup financing, and safe and trusted AI. Soket and ZenteiQ were selected under the Innovation Centre pillar, which covers foundation models, and twelve teams were shortlisted for indigenous models in total (PIB).
To build on the compute side, see our agentic AI tutorial for IndiaAI compute; for India's vendor scene, the agentic AI development company comparison; for upskilling, how to learn agentic AI in India and the best agentic AI course roundup.
Which model should you use?
- Choose Dhir 2B if you need an offline assistant now, work mostly in Indian languages, and have laptop or Jetson-class hardware.
- Choose Dhir 5B if code and maths matter and you can spare roughly 3 GB for the 4-bit file (ZenteiQ).
- Wait for EKA-1 if you are in banking, financial services or cybersecurity and want a larger model to fine-tune, but only commit once the weights and licence are published.
Limits to keep in mind: the Dhir context window is 16,384 tokens, which is short for long documents, and ZenteiQ has a 9B Dhir model listed as in development, so the lineup will shift (ZenteiQ).
FAQ
Q: What is EKA-1?
A: EKA-1 is a 26.3 billion parameter open, sparse hybrid model from Soket Labs, trained on about 7.5 trillion tokens across all 22 scheduled Indian languages plus English (PIB).
Q: Who made Dhir 2B and Dhir 5B?
A: Both were built by ZenteiQ AiTech Innovations, a company that began at IISc Bangalore, and released under Apache 2.0 through the AIKosh platform (ZenteiQ).
Q: Can Dhir run offline?
A: Yes. The 4-bit builds are 1.21 GB and 2.95 GB, and ZenteiQ reports 18.6 and 20.5 tokens per second on laptop graphics using llama.cpp (ZenteiQ).
Q: When will EKA-1 weights be public?
A: Soket's chief science officer Mayank Singh said an open release is planned within a month, with a waitlist for initial access, as reported by India Gazette.
Q: What is the IndiaAI Mission outlay?
A: The outlay is Rs 10,372 crore, approved by the Union Cabinet in March 2024 (PIB).
Top comments (0)