India wants its most sensitive systems on Indian AI. The first candidates arrived this week

Facts are attributed and linked to their sources. The analysis and opinions are the author's own.

What it is

On 3 November the IndiaAI Mission, part of the Ministry of Electronics and Information Technology, released three home-grown foundation models. EKA-1, from Soket Labs, has 26.3 billion parameters and was trained on about 7.5 trillion tokens covering all 22 scheduled Indian languages and English; the release describes it as an open, sparse hybrid model for maths, code, reasoning and agentic use. Soket's chief executive, Abhishek Upperwal, said it was built for India's banking, financial services and cybersecurity landscape. ZenteiQ AiTech released Dhīr 2B and Dhīr 5B, open-source assistants in English and 10 Indian languages that can run on-premise and offline, including on mobile phones. Both teams used subsidised national compute under the mission.

A day later Sarvam AI launched Prism, a decision model built on its Sarvam 105B base model for fast, structured choices such as routing customer-support queries, document analysis, security checks and grading AI agents. Sarvam says it answers in 44 milliseconds on an NVIDIA B200 GPU, and it released the weights under the Apache 2.0 licence so firms can train and run it on their own infrastructure. Its most useful claim is about confidence: when Prism said it was at least 90% sure, Sarvam says it was wrong 2.2% of the time in English and 4.4% in Indian languages, below its own predicted error rates.

Then, on 9 November, Business Standard reported that the government is likely to push for Indian AI models, LLMs and agents in critical government sectors such as defence and aerospace. An official said this could include asking critical departments, central and state, to avoid frontier models and agents built by non-Indian companies entirely. Officials told the paper that 20 models are approved under the ₹10,372 crore mission, being built by 12 startups, and that some domestic models have reached "up to 80 per cent" of the capacity of leading frontier models.

Why it matters

This is a policy direction, reported through unnamed officials, not a rule. But the direction matters to anyone who runs regulated systems in India, including banks, insurers and payment companies, because rules written for defence and government tend to become the reference point for other critical sectors.

A sovereign model is only useful in a critical system if it knows when it is unsure, and says so.

The "80 per cent" figure is the weakest number in the story. Eighty per cent on which tasks, measured how, and by whom? For a drafting assistant, 80% of frontier quality may be fine. For a model that routes a fraud alert or reads a loan document, the question is not the average score but the error rate on the cases that matter and whether the model flags its own doubt. That is why I think Sarvam's calibration claim is more interesting than its speed. A decision model that says "I am 90% sure" and is right about 97.8% of the time in English can be given a clear rule: act above the threshold, send the rest to a person. Those are Sarvam's own tests and need independent checking, especially the higher Indian-language error rate.

Open weights and offline deployment also change the procurement conversation. A model that runs inside the institution's own data centre, under a permissive licence, can be validated, versioned and frozen like any other model in a risk inventory. That is harder with a foreign frontier model that changes behind an API. The trade-off is capability: firms that switch for sovereignty reasons should measure the gap on their own tasks, not on a national average.

What I'd watch

• Whether the reported push becomes a written circular, which departments it covers, and whether financial regulators follow with their own guidance.
• Independent evaluations of EKA-1, Dhīr and Prism on Indian-language financial documents, with calibration reported alongside accuracy.
• Which of the remaining approved models are released in the next 12 to 15 months, and under what licences.