India wants artificial intelligence to speak more than English and not always depend on a distant data centre. On 3 November, the public IndiaAI Mission announced three models developed by two domestic companies: EKA-1 from Soket Labs, and Dhīr 2B and Dhīr 5B from ZenteiQ. Their launch is not yet proof of a new technological dominance. It is better understood as a strategic choice: to build tools adapted to India’s languages, public services and connectivity constraints.
Two model families, two deployment strategies
According to the Indian Ministry of Electronics and Information Technology, EKA-1 has 26.3 billion parameters and was trained on about 7.5 trillion tokens. The ministry describes it as a sparse hybrid model intended for mathematics, coding, reasoning and so-called agentic uses. It is said to support all 22 languages listed in India’s Constitution, alongside English. Soket says it has tailored the model to banking, financial services and cybersecurity.
Dhīr follows a different path. ZenteiQ’s 2B and 5B versions are presented as open-source multilingual assistants that can be deployed on premises and operate offline. The announcement says they support English and ten Indian languages. Proposed uses include searching documents and automating citizen services. The company says it is targeting sensitive or poorly connected environments, including government, industry, education and defence.
These specifications are claims made by the developers and government, not independently verified results in the announcement. The gap between a persuasive demonstration and a dependable tool will be measured through reproducible tests: quality in each language, robustness to local phrasing, security, operating costs and performance on the intended devices. “Multilingual” does not automatically mean equal mastery of every language, regional variant or spoken usage.
Language as economic infrastructure
For a country of more than a billion people, access to interfaces in the languages used at home, at work and in public services is an inclusion issue as well as a market opportunity. An assistant able to search documents in Hindi, Tamil or Bengali could reduce friction in education, administration or customer support. Its usefulness, however, will depend on practical details: available interfaces, data quality, speech support, accessibility for people less comfortable with writing and the ability to switch languages in the middle of a conversation.
Offline operation, if confirmed in real deployments, also changes the equation. An organisation can keep documents on its own systems and continue using an assistant when connectivity is unstable. That could matter for rural institutions, businesses subject to confidentiality rules or services that do not want to send data to a cloud provider. In return, local execution requires hardware, maintenance, security updates and a team able to monitor outputs.
Sovereignty is more than a national label
The government presents these models as a step towards strategic autonomy. The idea needs to be broken down. A model developed domestically can widen a country’s choices, but it does not by itself guarantee control over the whole supply chain: chips, energy, training data, software, hosting, maintenance and governance. Dependence may simply shift if essential components remain imported or developers publish too little information for independent assessment.
Public computing infrastructure is therefore a central part of the announcement. IndiaAI says the teams received subsidised access to national computing capacity as well as support for research expenses. That approach can help companies that could not finance large-scale training alone. It also raises an industrial-policy question: how should these resources be shared among teams, how can selection criteria remain transparent, and how should public returns be measured beyond a launch event?
Openness, safety and evidence of usefulness
ZenteiQ calls Dhīr open source, but that label should be assessed against what is actually published: model weights, code, licences, data or evaluation documentation. The ministry’s announcement does not detail every condition or provide a precise availability timetable. Useful openness for researchers and companies requires clear rights, practical documentation and channels for reporting defects. In critical environments, it also needs access controls and a process for fixes.
Models that may handle administrative requests or business documents need testing for fabricated answers, bias across languages and information leaks. Users should know when an answer is uncertain, which documentary sources were consulted and how to challenge an automated decision. Such safeguards do not diminish the appeal of a locally run model; they are part of the trust required for adoption.
Real-world use is the next measure
The launch of EKA-1 and Dhīr gives India a showcase for its AI policy and a starting point for discussion about models adapted to its linguistic realities. Success will depend less on announced parameter counts than on actual availability, comparative performance, costs and whether smaller organisations can deploy the systems. It will also depend on whether researchers, local authorities and speakers of the languages involved can help identify and correct blind spots.
If the models deliver, their relevance will extend beyond India. Many countries are looking for tools that work across languages, protect sensitive data and remain usable when cloud services are not an option. But ambition will need to be matched by reproducibility: public test sets, language-by-language measurements, clear limits and field reports can distinguish a durable industrial policy from a headline announcement. For now, India has introduced candidates and a support infrastructure. Their usefulness still has to be demonstrated with users.
The distinction matters in a crowded global AI market. An application that works well in a laboratory may struggle with dialects, mixed-language speech, low-cost devices or the messy records used by public offices. Pilot projects should therefore record not only speed and accuracy, but also failure rates, correction time, privacy incidents and the cost of keeping systems up to date. Publishing such evidence would help buyers compare local models with cloud alternatives on more than patriotic appeal.
India’s approach also illustrates how governments can shape markets without dictating every product. Shared compute, research grants and access to public datasets can lower entry barriers; procurement rules can create early demand. Yet public support is most credible when it is paired with open evaluation and clear terms for commercial and public use. If that balance holds, the launch may become a foundation for a wider ecosystem of language technology, rather than a one-off national showcase.
