India’s domestic large language model grant process has moved into its final stage, but the public record still stops short of the documents that would explain what was evaluated. A Techshots report on September 24 said the funding evaluations discussed in Vaishnaw’s September committee meeting were nearing completion. The same report did not publish the scorecards, the model list, the grant amounts, or the names of the evaluators. That absence is the story now.
The National e-Governance Division, or NeGD, is the administrative layer handling the domestic model grants. Its evaluators are not deciding whether a demo is impressive. They are deciding whether proposals meet the conditions that make public money useful: whether a model can be trained and served at a stated cost, whether it handles Indic languages beyond a token demo, whether training data can be traced, and whether safety claims can be tested by someone outside the applicant’s team.
What was evaluated?
On the available record, the evaluation covers proposals for domestic AI large language models under the government’s funding push. That is a broad category. It can include foundation models, domain models, and platforms that sit on top of existing open-weight systems. The Techshots report describes the process as a funding evaluation, not a procurement award. That distinction matters. A grant decision asks whether a team should receive public support to build something. A procurement decision asks which vendor should supply a service to the government.
The final stage, if it is genuinely final, should produce more than a shortlist. It should produce a defensible record: the evaluation rubric, the benchmark set, the compute assumptions, the language coverage tests, and the conflict-of-interest declarations. None of that has been published in the September 24 report. Until it is, the strongest claim is procedural: the process is closing. The substantive claim that the best domestic models have been identified remains unverified.
The CoRover shortlist is a different track
On September 26, Moneycontrol reported that MeitY had narrowed the field for an agentic AI platform to CoRover and Kyndryl, with DigiLocker as the first use case and a four-month build timeline. That is a separate procurement. It is about an agentic AI platform for a government service, not about grants for domestic foundation models. The two tracks may share a ministry, and they may share some vocabulary, but they do not share a decision memo.
Conflating them would be a mistake. A company that qualifies for an agentic platform pilot is not automatically a domestic model grantee. A model team that receives a grant is not automatically in line for a DigiLocker deployment. The September 26 shortlist answers a procurement question: which vendors can build an agentic layer for a specific first use case? The September 24 evaluation answers a funding question: which model proposals deserve public support. Readers should keep the two files separate.
What MeitY requires this month
The ministry’s immediate requirement is not another vision statement. It is paperwork that can survive scrutiny. If the evaluation is complete, MeitY should be able to publish a final list or, at minimum, a status note that explains what remains. The note should say whether grants are conditional, whether compute credits are part of the package, and whether recipients must meet milestone-based tests before disbursement.
It should also say how evaluators handled the obvious risks. Domestic model grants can fund genuine capability, or they can fund polished interfaces around imported weights. The difference is not rhetorical. It shows up in training runs, data provenance, evaluation harnesses, and inference costs. A credible grant process would require applicants to document those things, not just promise them.
Why the delay matters
India’s AI policy has a credibility problem when announcements arrive before artifacts. The country has compute schemes, mission documents, and pilot programs. What it has less of is a public evaluation trail. Every week that the model grant list stays unpublished, the ecosystem fills the gap with speculation. Startups guess at the criteria. Researchers cannot check the benchmarks. State governments do not know which models they can safely procure.
The September 24 Techshots report is useful because it marks a stage, not a conclusion. It tells us the Vaishnaw-led review is close to finishing. It does not tell us what finished means. That is not a reason to dismiss the process. It is a reason to ask for the documents that turn a stage into a decision.
What to watch
Three things will show whether this is a real evaluation or a press cycle. First, the final list: names, model families, and grant amounts. Second, the rubric: the benchmarks, language tests, safety checks, and compute assumptions used to rank proposals. Third, the separation: a clear statement that the agentic AI procurement and the domestic model grants are distinct tracks with distinct winners.
Until those appear, the responsible read is narrow. The evaluation is nearing completion. The CoRover-Kyndryl agentic shortlist is a different story. And the domestic model grant record remains, for now, a final stage without a final scorecard.