Soket AI Labs, one of the 12 startups selected under the government’s IndiaAI Mission, is building a 120-billion-parameter artificial intelligence model aimed at coding, with founder Abhishek Upperwal saying the company is seeing early interest from domestic IT services firms as well as government stakeholders.
Speaking to FE, Upperwal said the new model will build on the company’s earlier small language model, Pragna-1B, which was developed as an experiment to see whether foundational AI models could be built from India with support for Indic languages.
From Small Experiments
“When we started almost two years ago, Pragna-1B was essentially an attempt to see whether we could build a foundational model from India in Indic languages,” Upperwal said. “Now we are building a 120-billion parameter model which will support all Indic languages, including the 22 scheduled languages.”
The company intends the model to serve as a base platform for applications built in India and across emerging markets. Upperwal said Soket AI is also open-sourcing parts of its datasets to encourage wider development.
A key use case the startup is targeting is coding, particularly for the domestic large IT services industry. According to Upperwal, several firms are exploring domestic AI models as an alternative to global offerings.
“We are seeing interest from IT companies for the AI model we’re building for coding,” he said. “Many of them want domestic models so that their data remains within the country and operates within defined guardrails.”
Catering to Strategic Sectors
Upperwal added that demand is currently strongest from government departments and public-sector entities, but private-sector interest is also rising. Soket AI is in discussions with stakeholders in sectors such as defence, agriculture and financial services for potential deployments.
“There are deals in the works which we will announce soon,” he said, adding that AI models can have applications in areas such as strategic analysis and real-time monitoring in defence.
For now, the company is focusing on building a strong text-based foundational model before expanding to multi-modal capabilities.
“We are working on the brain first,” Upperwal said. “Once that is built, we can add the eyes and ears later. Multi-modality can be added later, but the base model needs to be strong.”
Upperwal said the company expects its coding model to achieve performance comparable to tools such as Anthropic’s Claude Code within about a year, supported by investment in specialised datasets.
Funding for AI startups in India is still modest compared with the US, he acknowledged, but early-stage interest is improving. Soket AI plans to raise a seed round soon.
“We cannot expect funding at the scale of the US,” Upperwal said. “But seed-stage investment in AI startups in India will grow. Traditional businesses are also showing interest in investing.”
He also said infrastructure under the IndiaAI Mission is gradually improving despite early delays in the rollout of GPUs for startups.
“This is the first time the country is deploying AI servers at this scale,” Upperwal said. “We are working with the infrastructure teams to optimise usage. There are around 6,000–8,000 usable H100 GPUs now and that could go beyond 10,000–15,000 within a year.”