◆ Live signal · AI Infrastructure

India releases EKA-1 and Dhir foundation models under IndiaAI: Indian models move towards deployment, but the chips beneath them are still imported

Signal in brief
  • MeitY released Soket's 26.3-billion-parameter EKA-1 and ZenteiQ's Dhir 2B and 5B under the IndiaAI Mission on 3 November 2026.
  • Dhir is open-source (Apache 2.0), pretrained from scratch and reported to run on laptop-class and embedded hardware; it was trained on Google TPUs and Nvidia GPUs in Mumbai.
  • EKA-1 has no public benchmarks, weights or licence yet, and all Dhir performance figures are company-reported.
Key claims
  • MeitY released Soket's 26.3-billion-parameter EKA-1 and ZenteiQ's Dhir 2B and 5B under the IndiaAI Mission on 3 November 2026.
  • Dhir is open-source (Apache 2.0), pretrained from scratch and reported to run on laptop-class and embedded hardware; it was trained on Google TPUs and Nvidia GPUs in Mumbai.
  • EKA-1 has no public benchmarks, weights or licence yet, and all Dhir performance figures are company-reported.
Primary sources
  • Ministry of Electronics and IT / PIB (3 Nov 2026): https://www.pib.gov.in/PressReleseDetailm.aspx?PRID=2332315&lang=1&reg=3
  • ZenteiQ, Dhir technical page (accessed 5 Nov 2026): https://www.zenteiq.ai/dhir/
  • Soket AI, Project EKA (accessed 5 Nov 2026): https://soket.ai/project-eka

The Ministry of Electronics and IT (MeitY) released two sets of indigenous foundation models at a press event in New Delhi on 3 November 2026, in the presence of MeitY Secretary S. Krishnan, according to the PIB release: EKA-1 from Soket Labs Technology and Research, and Dhir 2B and Dhir 5B from ZenteiQ AiTech Innovations. Both companies were selected under the IndiaAI Innovation Centre's foundation-models pillar, and the government says subsidised access to national AI compute and support for research expenses enabled the work.

EKA-1 (Soket Labs)

  • Size: 26.3 billion parameters, trained on about 7.5 trillion tokens (PIB).
  • Languages: all 22 scheduled Indian languages and English.
  • Design: described as an open, sparse hybrid language model for maths, code, reasoning and agentic use.
  • Positioning: Soket CEO Abhishek Upperwal says it is tailored to India's banking, financial services and cybersecurity needs. That is the company's framing.
  • Caution: Soket's own Project EKA page does not mention EKA-1's figures and instead describes a 120-billion-plus-parameter model in training on 1,536 H100 GPUs, which appears to be a separate, larger effort. No benchmark results, weights or licence terms for EKA-1 were found.

Dhir 2B and 5B (ZenteiQ)

  • Dhir 2B: a dense model of 2.03 billion parameters.
  • Dhir 5B: a mixture-of-experts model with 5.09 billion parameters in total, of which 1.79 billion are active per token (2 of 12 experts).
  • Both: 16,384-token context, English plus ten Indian languages (Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi and Odia), pretrained from scratch rather than adapted from another developer's model, released under the Apache 2.0 open-source licence.
  • File sizes: 1.21 GB and 2.95 GB in 4-bit form; 4.06 GB and 10.19 GB at 16-bit.
  • Training, per ZenteiQ: Google Cloud TPU v6e in Mumbai, with 2,048 TPU v6e chips and 64 Nvidia H100 GPUs, over 6 trillion pretraining tokens per model, a 131k-entry in-house tokenizer, and data and checkpoints kept in the same Indian region.
  • Edge speed, per ZenteiQ (4-bit, single stream, llama.cpp; 2B / 5B): Intel Iris Xe integrated graphics 18.6 / 20.5 tokens per second; Nvidia Jetson Orin NX 34.1 / 39.9; Raspberry Pi 4 CPU 2.3 / 2.8.

All Dhir specifications, benchmark scores and speeds are company-reported and have not been independently verified.

Why it matters

The IndiaAI Mission, approved on 7 March 2024 with seven pillars from compute to datasets and safe AI, is now producing released models rather than selected proposals. The two releases take different routes. EKA-1 aims at a mid-size general model with full Indian-language coverage. Dhir aims at small models that run offline on a laptop-class chip or an embedded board, which matters for factories, public services and sensitive sites that cannot send data to the cloud.

The release also shows the limit of model sovereignty. Dhir was built in India, on Indian-region infrastructure, but on Google TPUs and Nvidia GPUs. Owning the model is one layer; accelerators, memory, networking and data-centre hardware remain imported. That connects this signal to the compute tender in S-165 and the fibre build in S-171. Offline use also does not by itself make a model fit for defence work, which needs security accreditation and testing (see S-176).

What has not been disclosed

  • EKA-1 technical report, architecture detail, active-parameter count, weights, licence, hardware needs and benchmarks.
  • Independent tests of Dhir accuracy and speed, and its performance under many concurrent users.
  • Commercial deployments, customers and support arrangements for either model.
  • How much subsidised IndiaAI compute these projects used.

What to watch

Watch for EKA-1 weights and a model card, independent Indian-language benchmarks, first enterprise or government deployments, and whether Indian hardware makers begin supplying edge devices that run models like Dhir.

Share this

Track the systems we watch

Signals, reports and briefings on India’s industrial transformation.

Subscribe →
Was this signal useful?
Keep reading

Related research