IndiaAI Releases EKA-1, a 26.3-Billion-Parameter Model Covering All 22 Scheduled Languages, and Dhīr 2B and 5B, Open Models Built to Run Offline

On 3 November 2026 the IndiaAI Mission released two sets of home-built AI models. One is EKA-1, from Soket Labs of Bengaluru, a model of 26.3 billion parameters that the Press Information Bureau says spans all 22 scheduled Indian languages along with English. The other is Dhīr 2B and Dhīr 5B, from ZenteiQ, small models for English and ten Indian languages that are built to run on a laptop or an edge device with no internet connection. Parameters are the adjustable numbers inside a model, a rough measure of its size.

Unveiled in New Delhi with the IT Secretary

The PIB release says the models were unveiled at a press event in New Delhi attended by S Krishnan, Secretary of the Ministry of Electronics and IT, and officials of the ministry and IndiaAI, the mission’s implementing division. Krishnan said such models strengthen India’s “technological and strategic autonomy in Artificial Intelligence”. He added that platforms such as AIKosh, the mission’s datasets and models library, can make them available to researchers, students, innovators and the wider public.

Sudeep Shrivastava, chief operating officer of IndiaAI, said the models show progress from lightweight solutions for everyday devices to heavier models for complex enterprise work. Both companies are among the twelve teams selected under the mission’s foundation-model pillar, and the release says subsidised national AI compute and support for related research costs made the models possible.

EKA-1, a model for every scheduled language

PIB gives EKA-1 about 7.5 trillion tokens of training text, a token being a word or a piece of one. It describes the model as an open, sparse hybrid language model built for mathematics, code, reasoning and agentic use, meaning software that carries out a task in several steps. The 22 scheduled languages are those listed in the Eighth Schedule of the Constitution, which our explainer sets out.

The model’s page on AIKosh adds detail. It says Soket built EKA-1 from scratch with IndiaAI support and calls it the first release from Project EKA, Soket’s research programme. The architecture, which Soket names RALSNet, combines attention, recurrent and expert layers in a sparse design, so that only part of the network works on each token. Soket reports inference more than three times faster than comparably sized models, and a context window of 64,000 tokens, the amount of text the model can read at once. The page says internal evaluations show performance on par with comparably sized models, with methods and results to follow in a technical report.

Soket chief executive Abhishek Upperwal said in the PIB release that the model is tailored to banking, financial services and cybersecurity in India. ANI reported that the model is designed to run on private servers, private cloud and air-gapped systems, which have no link to the outside internet. It quoted Soket co-founder Mayank Singh as saying data preparation, training and evaluation were done entirely in India.

Sources count languages differently. PIB says 22 scheduled languages plus English. The AIKosh page says more than 100 Indian, Global South and programming languages, and Soket’s home page says 60 or more, so those counts include languages outside India’s schedule and programming languages.

Dhīr 2B and 5B, models that work without a network

ZenteiQ was founded in 2022 by Sashikumaar Ganesan, a professor at the Indian Institute of Science in Bengaluru, according to The Hindu BusinessLine. PIB names the models as open-source, multilingual assistants for secure, on-premise, offline-capable use, with document-grounded automation, knowledge retrieval and citizen or enterprise services in defence, government, industry and education. It points to low-connectivity and sensitive settings.

ZenteiQ’s model page gives the numbers. Dhīr 2B is a dense model with 2.03 billion parameters. Dhīr 5B is a mixture-of-experts model, 5.09 billion parameters in all with 1.79 billion active for each token, using 2 of 12 experts. The company’s technical report abstract says both were pretrained from scratch on more than six trillion tokens, with an in-house tokenizer of 131,000 entries, and that every training stage ran on Google Cloud TPU v6e chips in India. The ten Indian languages are Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi and Odia.

On size, the abstract says Dhīr 2B, with its weights compressed to 4 bits, is a 1.21 GB file that runs fully offline on a laptop’s integrated graphics and on an NVIDIA Jetson module, a small board for edge devices. The page lists speeds measured with the llama.cpp program: 18.6 tokens a second for Dhīr 2B and 20.5 for Dhīr 5B on Intel Iris Xe laptop graphics, and 2.3 and 2.8 on a Raspberry Pi 4 processor. The weights carry an Apache 2.0 licence, which allows free use and changes.

At the release Ganesan said the models run on GPUs and TPUs and, being light, can be used on a mobile phone with no cloud link, keeping data on the device. He told ANI the company is building a free app for Android and iOS.

What the mission paid for

The Rajya Sabha reply of 13 February 2026, reproduced in a PIB release, lists the support given to each of the twelve teams. Soket AI received Rs 177.08 crore, of which Rs 162.47 crore was compute and Rs 14.61 crore other costs. ZenteiQ received Rs 206.49 crore, of which Rs 165.19 crore was compute and Rs 41.30 crore other costs. Together that is Rs 383.57 crore, by tuput’s addition.

PIB announced Soket’s selection on 30 May 2025 with a plan for an open-source model of 120 billion parameters. EKA-1 is a smaller first release, and Soket’s home page lists a model of more than 120 billion parameters as in training. The mission’s compute pricing, its other funded teams and Sarvam’s models are covered in an earlier tuput piece on the IndiaAI Mission, and the language-translation side is in our piece on India building AI in its own languages.

What is still to be shown

As of 5 November 2026, no outside tester had published benchmark scores for the three models that tuput could find, so the performance figures here are the developers’. ZenteiQ says Dhīr 2B is first in its class on 13 of 15 Indic benchmarks, and that both models were compared with 24 open models of one to seven billion parameters across 41 benchmarks. It says the full tables are in its technical report, which its publications page lists as coming soon. The same page lists limits: recall of facts without a supporting source is the weakest area, so it advises retrieval from authoritative sources and human review for law, finance, agriculture, medicine or history. It says neither model has a reasoning mode and that both handle text only.

On access, the AIKosh page for EKA-1-26B showed the Apache 2.0 licence, a Restricted setting, under which downloads need the owner’s approval, and no downloads. Soket’s home page lists its EKA-26B code and reasoning model as coming soon. ANI reported that Mayank Singh said Soket plans to release the weights openly within a month, starting with a waitlist, and that the Dhīr weights are already on AIKosh. ZenteiQ also told ANI it plans a 9-billion-parameter model for December.