DRONA 2.0: A Staff College in the Nilgiris Builds on Sarvam

The Defence Services Staff College at Wellington in the Nilgiris has launched DRONA 2.0, a multimodal generative AI system built on Sarvam's indigenous 105B model alongside open-source LLMs.

It was launched on 18 August by Lt Gen Manish Erry, Commandant DSSC, with Dr Vivek Raghavan, co-founder of Sarvam AI, present. An MoU was signed alongside it covering collaboration on DRONA GPT and further AI applications. The details come via Headquarters Integrated Defence Staff.

Two days earlier, a benchmarking analysis had India's best model sitting roughly where the global frontier stood two years ago. This is what deploying that model looks like in practice.

What DRONA 2.0 does

It runs on a dedicated GPU server at DSSC, and the capability list is broad for an institutional deployment:

• Reasoning and document inferencing
• Image recognition and generation
• Speech-to-text and text-to-speech
• Translation and transliteration
• Podcast and presentation generation
• Data analytics tooling

The system was trained on DSSC's own academic repository, which is the design decision that matters. It's positioned as a high-precision resource for staff college officers and trainees rather than a general assistant — grounded in the institution's own material rather than expected to know it already.

This follows DRONA 1.0, launched last year, an indigenously developed system aimed at optimising resource management and training.

Why the Sarvam choice is the story

A staff college could have deployed a frontier model through an API and had better raw benchmark performance on day one. It didn't.

A defence training institution is not solving for the highest coding benchmark. It is solving for where the weights sit, who can see the queries, whether the system keeps working if a commercial relationship or an export control changes, and whether its own corpus ever leaves its infrastructure.

An open-weights Indian model on a GPU server inside the campus answers all four in a way an API call cannot, at any benchmark score.

The translation and transliteration capability points the same direction. That is the specific capability global leaderboards do not test — none of the standard benchmark suites measure an Indian language — and it is precisely the capability Sarvam was built for.

The broader pattern

HQ IDS frames this as part of wider efforts to infuse AI into military training.

The interesting signal for anyone tracking Indian AI is the shape of the adoption. Not a pilot, not a proof of concept, and not a frontier model wrapped in a procurement contract — an indigenous open-weights model, fine-tuned on institutional data, running on owned hardware, with an MoU for continued joint development.

That combination is a repeatable template, and it is available to any Indian institution with sensitive data and a reason to keep it on premises. Which is to say: most of them.

The benchmark gap is real. It also isn't the only variable in a deployment decision, and DRONA 2.0 is a useful reminder of the other ones.

Source: The Times of India — DSSC launches multimodal AI system using Sarvam to boost defence research