Infosys, SAP Labs and IBM are seeing growing interest in small language models (SLMs) rather than large language models (LLMs) as enterprises look for models that can deliver targeted business solutions while addressing challenges such as computational demands, cost, and data security.
Infosys launched two SLMs about a month ago using the NVIDIA AI stack and targeting banking and IT-specific applications. It is also seeing increasing interest from enterprises worldwide for tailored SLMs. Balakrishna DR, EVP, global services head, AI & industry verticals, said: A lot of our customers are saying, ‘Can you actually develop it for my context?’ So, all of these are opportunities for us to help the entire world adopt AI,” he said. SLMs, often streamlined versions of LLMs, are designed with fewer parameters and simpler architectures, making them more efficient and easier to deploy.
Vishal Chahal, V-P of IBM India Software Lab, said SLMs like IBM’s 8B or 3B Granite models are purpose-built with fewer parameters requiring less data, offering significant cost savings. “These models are ideal for deployment on smartphones and edge devices, contributing to sustainability,” Chahal said.
He added that SLMs also simplify risk management by reducing issues such as bias and hallucination due to their smaller and more focused datasets. “Smaller models are typically trained on highly curated data with controls and filters to ensure data accuracy, completeness, and provenance,” he said.
Sanjeev Menon, co-founder and head of product at E42.ai, said when it comes to enterprise proprietary data, SLMs fare much better because of the model size and the ability to fine-tune with limited data. “This helps reduce hallucinations and ensures generations are derived solely from enterprise data,” he said.
SAP Labs, too, seeing significant traction for its custom AI and agentic AI solutions, particularly in emerging markets. Milesh J, head of strategy & operations, said the company’s business AI solution is for all global customers. “But when we talk about custom AI, it’s about personalising it and making it very specific to a target market in India,” he said.
SLMs are machine learning algorithms trained on more specific datasets, making them more agile and adaptable to targeted tasks compared with LLMs, which are general-purpose models with billions of parameters. LLMs are often expensive to train and maintain, making them less ideal for edge or domain-specific applications.
Ganesh Gopalan, co-founder & CEO of Gnani.ai, noted that SLMs are particularly beneficial for domains such as BFSI, customer service, and fraud detection. “SLMs provide faster performance, cost-effectiveness, and seamless integration, making them ideal for localised or domain-specific use cases. By leveraging domain-specific training, SLMs and advanced voice-to-voice models reduce biases and enhance relevance in customer interactions,” Gopalan said.
The customisation capabilities of SLMs are a significant factor driving their adoption. As SAP Labs’ Milesh said, “We’re also introducing the agentic framework with multi-agents that can handle complex business functions.” These agentic AI systems can perform complex tasks with minimal human intervention, making them a preferred choice for many enterprises.
The sustainability aspect of SLMs further adds to their appeal. “Being lightweight and less computationally intensive, SLMs are ideal for deployment on edge devices,” IBM’s Chahal said.
