Soket Labs and ZenteiQ release Indian AI models that run offline

Two startups backed by India’s government, Soket Labs and ZenteiQ, released three foundation models on Tuesday, 3 November, under the IndiaAI Mission. Soket’s EKA-1 and ZenteiQ’s Dhīr 2B and Dhīr 5B are built to work in airgapped environments, meaning systems fully cut off from the internet and outside networks. The models were unveiled at a press briefing in New Delhi attended by IT Ministry secretary S Krishnan and IndiaAI Mission COO Sudeep Shrivastava.

The pitch is sovereignty and offline use rather than leaderboard position: the IT ministry framed the launch as a step towards homegrown AI capabilities, and the announcement did not include benchmark results, deployment details or pricing.

What Soket Labs and ZenteiQ released under the IndiaAI Mission

The IndiaAI Mission, run by the Ministry of Electronics and IT (MeitY), released EKA-1 from Soket Labs Technology and Research, and Dhīr 2B and Dhīr 5B from ZenteiQ AiTech Innovations. Both companies were selected under the mission’s IndiaAI Innovation Centre (Foundation Models) pillar, which aims to build indigenous AI capabilities and large language models.

The two releases target different jobs. Krishnan said Soket’s model is aimed at maths and coding, while ZenteiQ’s is aimed at science and research. Inc42 summarised the split as EKA-1 targeting mathematics, coding, reasoning and agentic use cases, and the Dhīr models focusing on smaller multilingual assistants that can operate with limited connectivity.

Spec	EKA-1 (Soket Labs)	Dhīr 2B / Dhīr 5B (ZenteiQ)
Size	26.3 billion parameters	2 billion and 5 billion parameters
Training data	About 7.5 trillion tokens	Not stated
Languages	All 22 scheduled Indian languages plus English	English plus 10 Indian languages
Openness	Described as “open” by Soket	Open-source
Stated focus	Maths, code, reasoning, agentic use; banking, financial services and cybersecurity	On-premise multilingual assistants, document-grounded automation, knowledge retrieval
Hardware	Certain inferencing on a local device such as a smartphone	GPUs and TPUs; phones with at least 6 GB of memory

EKA-1: a 26.3B sparse hybrid model for maths, code and finance

EKA-1 has 26.3 billion parameters and was trained on about 7.5 trillion tokens, the units of text a model learns from. It covers all 22 scheduled Indian languages along with English. Soket describes it as an “open, sparse hybrid language model” designed for math, code, reasoning and agentic use; the coverage does not explain the architecture behind that description.

Soket co-founder and CEO Abhishek Upperwal positioned the model around specific Indian sectors:

“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. It is designed to address sector-specific requirements, understand India’s evolving digital ecosystem, and deliver solutions aligned with the country’s needs and priorities.”

That quote comes from the ministry’s statement as reported by Inc42. PTI adds that EKA-1 can be used without a mobile data or WiFi connection, because it can do certain inferencing on a local device such as a smartphone. Upperwal also said safeguards are in place, including protections against content on suicide and self-harm.

The launched model differs from what Soket described in early 2025. In April 2025, Upperwal told Analytics India Magazine that the company had proposed a 120-billion-parameter open-source Indic LLM trained on 2 trillion tokens under its “EKA Project”, on a 12-month timeline, with plans for multiple small models within six months. EKA-1 is smaller in parameters than that proposal but trained on more tokens. None of the launch coverage says whether the larger model is still planned.

Dhīr 2B and 5B: compact open-source models for phones and on-premise

ZenteiQ’s two models are open-source multilingual assistants supporting English and 10 Indian languages, designed for secure, on-premise and offline use in critical sectors. The ministry lists 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 CEO Professor Sashikumaar Ganesan told ANI:

“We have launched two small models. We call them compact models...what we have launched is a two-billion and then five-billion parameter model...”

(quoted by The Economic Times). His name is spelled Sashikumaar in that report and Sashikumar in others.

In the ministry statement, he said the models can run across different computing platforms, including GPUs and TPUs, and that their lightweight architecture allows on-device deployment on mobile phones without dependence on cloud connectivity, keeping data on users’ devices. At the briefing, he said the model can run on a mobile phone with at least 6 GB of memory.

How airgapped deployment changes the picture

Ganesan contrasted the setup with popular chatbots. With a service like ChatGPT, he said, the “inferencing”, the step where the model produces an answer, happens on a server at a remote location; in an airgapped setup it happens locally. “No data will go outside,” he said, adding that the model needs no network connection. He named defence and cybersecurity as possible users.

Krishnan made a similar case for small models generally, saying they have many uses, including cybersecurity, and can be deployed in airgapped environments. He also said indigenous models strengthen India’s technological and strategic autonomy in AI and open opportunities for productivity across government, industry, research and education.

The compute behind the models

The government said access to subsidised national AI compute infrastructure, along with support for ancillary research expenses, enabled the development of the models. Ganesan said ZenteiQ trained on Tensor Processing Units (TPUs) because GPUs were scarce. He said TPUs are not inferior to GPUs, and that getting GPUs would have meant waiting another three to four months.

For context, MeitY had onboarded 10 GPU-as-a-service providers in 2025 to build centralised compute for the mission, aiming for GPU access below $1 per hour, according to Analytics India Magazine’s report at the time. The ZenteiQ account suggests that capacity was still tight enough to push at least one team to other hardware.

What the launch does not tell builders yet

The launch is light on the details developers need to evaluate a model. Inc42 notes that the announcement did not provide benchmark results, deployment details or pricing information. There are no published scores, even from the makers themselves, so claims about maths, coding or science ability cannot be checked yet.

Other gaps in the coverage:

• Licence terms. Dhīr is called open-source and EKA-1 “open”, but no source names a licence or says whether commercial use is allowed.
• Where to get the weights. No download location is given. Krishnan said platforms such as AIKosh can make the models accessible to researchers, students, innovators and the wider public, which describes a possibility rather than a confirmed release channel.
• Context length, speed and memory needs for EKA-1. The 6 GB phone figure was given by ZenteiQ for its own models; no hardware requirement is reported for EKA-1.
• The Dhīr language list. Sources say 10 Indian languages but do not name them.

Upperwal’s mention of self-harm protections is the only safety detail reported. For models meant to run on disconnected devices, where the maker cannot update or monitor behaviour after deployment, how those safeguards work and how they are tested will matter.

What it means for developers building with Indian languages

For teams working in regulated or sensitive settings, such as banks, defence suppliers and government departments, the practical appeal is local inference: data stays on the device or on-premise, and nothing depends on a cloud API. A 2B or 5B model that runs on a 6 GB phone could suit field apps in low-connectivity areas, which the ministry names as a target.

EKA-1’s coverage of all 22 scheduled languages is broader than Dhīr’s 10, so builders who need less widely supported Indian languages may look there first, while Dhīr’s smaller sizes fit tighter hardware. Until benchmarks and licences appear, both are candidates to test, not drop-in choices.

The launch landed the same week Sarvam AI released Prism, an open-weight decision model built on Sarvam 105B; Unicorn API covered that release in Sarvam AI open-sources Prism, a model that picks answers in 44 ms. Together they show Indian model makers aiming at different deployment needs rather than one general chatbot.

What to watch next

• A 9B Dhīr model. Ganesan said ZenteiQ plans to launch a 9-billion-parameter model.
• A free mobile app. He also said ZenteiQ’s mobile app will come to the Play Store and iOS and will be free.
• Benchmarks and model cards. Any published evaluation, first from the makers and then from independent testers, will be the first real test of the maths, code and science claims.
• Licences and distribution. Whether the weights appear on AIKosh or elsewhere, and under what terms.
• Soket’s larger model. Whether the 120B-parameter EKA Project model proposed in 2025 is still on the roadmap.