Shrinking AI: Companies build smaller-scale tools to provide tailored solutions

While these and other Indian companies go all in on AI, many are developing and customising small language models (SLMs) because off-the-shelf AI tools either fall short of their needs or are too wide-ranging for their functions, experts told Mint.

Aakash Educational Services Ltd, Physics Wallah, Healthify and Zoho are among those betting big on such tools, also known as internal foundational AI models. Aakash, which coaches students for professional entrance examinations, is developing AI models that are expected to fructify in 18-24 months, MD and CEO Deepak Mehrotra told Mint. 

Its focus is on assessment analytics, which refers to analysing test results, learner performance and skills evaluations. It also plans to use AI to create custom learning packages and career pathways.

“We’ve taken baby steps in this direction, but we’re already seeing the impact,” said Mehrotra.

The company has started rolling out the initial versions of its AI offerings through Invictus, the latest version of its preparation programme for the Joint Entrance Examinations for India’s engineering and architecture courses.

The next step: using proprietary data and partnering with external experts, including big tech companies, to develop SLMs. Mehrotra did not disclose further details on partnerships.

Economical option

Unlike large language models (LLMs) like ChatGPT, which use hundreds of billions or even trillions of parameters, SLMs are trained on significantly fewer parameters, ranging from a few million to a few billion. 

“SLMs offer a cost-effective alternative for GenAI (generative AI) development and deployment since many organisations don’t need broad capabilities,” said Anushree Verma, director analyst at tech consulting firm Gartner. 

SLMs are easier to control and fine-tune for specific use-cases, she added.

Companies with large proprietary data pools are leaning toward building SLMs to deliver targeted offerings, especially where existing LLMs fall short of their needs.

Some are building them over existing models, also known as LLM distillation. Through distillation, important functions and knowledge can be derived from a larger model and deployed in a smaller one. Organisations can tweak these smaller models with their own data to tailor them to specific use-cases.

Others are likely to give it a go from scratch. 

Cloud software startup Zoho Corp plans to make two AI-based internal foundational models public by the end of the year, Mint reported last week, citing group chief executive officer Shailesh Davey. Zoho confirmed to Mint that the small and medium models will be built from scratch.

Edtech relevance

Two of the four companies that affirmed their plans to build internal foundational models to Mint – Physics Wallah and Aakash – are from the edtech space, indicating the rising relevance of these tools in this segment.

Mint reported earlier that Physics Wallah plans to roll out its SLMs by end-2026. The firm is building models focused on physics, chemistry, mathematics and biology – areas where current LLMs have limited effectiveness.

However, the need for SLMs may vary across segments within edtech. Ankur Dhawan, chief product and technology officer at upGrad, said SLMs are most effective in structured and static curricula like K-12 education, “where conceptual learning is key.”

“In higher education, interdisciplinary learning and broader content, the requirements make LLMs more suitable,” he said. “The cost-effectiveness and latency of existing models are already strong and improving rapidly, driven by AI innovation. Given this, we don’t see an immediate need for SLMs in our domain.”

Most efficient

Dhawan added that developing and maintaining niche models would also require substantial ongoing investment, potentially slowing down innovation.

“Given the fast-paced advancements in AI, we believe that leveraging the best available models is the most efficient way to serve our learners and stay ahead of industry transformations,” he said.

While comparatively smaller than LLMs in scale, internal foundational models also incur a high cost and long timelines when made from scratch.

Aakash had committed at least ₹100 crore toward AI and digital tech last year. Mehrotra forecasts this investment will grow in the next two to three years even as it is attempting to build models frugally. 

In an interview earlier this week, Physics Wallah founder and chief executive officer Alakh Pandey said the company is making significant AI investments without overanalysing, noting that timing would be crucial in capturing the opportunity.

Physics Wallah did not share additional details because it is in the run-up period to its initial public offering.

Zoho and Healthify did not immediately say how much they are investing in developing alternative AI tools. 

To deal with the increasing cost of building AI models from scratch, companies are opting for distillation.

Healthify, a health and wellness startup, has developed SLMs that run on top of LLMs from OpenAI and Anthropic.

“We’re layering it with our own native data—lifestyle, diet patterns, locations, preferences and medical conditions. This creates a more personalised and intelligent tool,” said company CEO Tushar Vashisht.

The company sees value in SLMs primarily in training its AI assistant Ria and its coaching co-assistant to provide real-time conversational advice that feels human and empathetic. As a secondary function, the company’s SLM provides contextual nudges to users alongside personalised health insights in an effort to retain customers and drive engagement.

The company didn’t see the need to build from scratch.

“Building SLMs from scratch is resource-intensive and time-consuming. Using existing models helps us quickly integrate advanced AI capabilities without heavy upfront costs or long lead times,” said Healthify’s chief technology officer, Abhijit Khasnis.