How domain-specific SLMs are reshaping enterprise AI
As AI adoption matures in the enterprise, the spotlight is shifting from massive, general-purpose models to smaller, bespoke, specialised ones built for precision and efficiency. Small Language Models (SLMs) are at the heart of this evolution. Unlike large foundational models, SLMs deliver domain-specific intelligence that fits within real-world constraints, ice compute limits, regulatory requirements, and the growing need for transparency.
By 2025, organisations across industries began deploying SLMs at scale. The move was driven by a simple truth: smaller models, tuned for specific tasks or sectors, often outperform their larger counterparts in speed, security, and accuracy. Their lightweight design enables on-premises or edge deployment, reducing latency and minimising reliance on cloud infrastructure. This approach also aligns with global data residency and AI assurance mandates, where keeping data within enterprise boundaries is no longer optional.
Most SLMs operate with fewer than 10 billion parameters, making them ideal for commodity hardware or edge devices. Some models have fewer than a billion parameters, while others, like Microsoft’s Phi-3 (3.8B) and Google’s Gemma (2B and 7B), sit comfortably in the multi-billion range. Thanks to advances in techniques, distillation, quantisation, and instruction tuning, these models deliver strong performance despite their size. When deployed with retrieval-augmented generation (RAG), they can tap into external knowledge bases for highly contextual responses. This blend of compactness and contextuality further allows enterprises to unlock AI value without ballooning infrastructure costs.
Fine-tuning is where SLMs shine. Trained in domain-specific datasets, i.e. clinical notes in healthcare, legal contracts in compliance or purchase orders in supply chain, makes them more relevant. They consistently outperform generic large models on specialised tasks. For example, an energy company using a fine-tuned SLM for equipment failure prediction reduced downtime significantly. The SLM processed SCADA logs locally, which reduced latency, thus enabling near-real-time anomaly detection in sites with limited connectivity. Similar setups are now common in industrial environments, where SLMs power diagnostics, maintenance alerts, and local decision-making without sending sensitive data offsite.
SLMs also support sustainability goals. Research from European institutes shows that training a 6-billion-parameter model emits as much CO₂ as a transatlantic flight. SLMs, by contrast, consume far less energy during training and fine-tuning. This is helping enterprises meet carbon neutrality targets and comply with emerging Green AI standards in regions like the EU.
From a governance standpoint, SLMs offer clear advantages. Smaller models are easier to debug, interpret, and document, which are critical in sectors like finance and healthcare where decisions must be transparent and auditable. Since they are often developed in-house and focused towards a sector, enterprises can have better control over them. Organisations adopting Responsible AI frameworks find SLMs inherently more compliant. They can embed custom ethical guardrails and validate against domain-specific fairness and safety metrics. This is far harder to achieve with opaque, general-purpose models trained on uncurated public data.
The rise of SLMs is also fueled by open ecosystems. Open-source communities have released high-performing, commercially viable SLM architectures under permissive licenses. Enterprises are contributing back with benchmarks, training recipes, and alignment datasets. This is driving innovation while reducing dependence on proprietary platforms and vendor lock-in.
Looking ahead, hybrid architecture will dominate. Large models will still handle complex reasoning, but most operational tasks like document processing, customer interactions, and anomaly detection will be managed by SLMs. In orchestrated agentic systems, SLMs will act as domain specialists, delivering fast, traceable decisions aligned with enterprise goals.
The shift to domain-specific SLMs marks a turning point in enterprise AI strategy. It’s a move away from one-size-fits-all intelligence toward fit-for-purpose AI that is efficient, interpretable, and secure. In a world shaped by regulatory pressures and infrastructure constraints, small language models will be a strategic necessity..
Balakrishna D.R.
Balakrishna D. R is Executive Vice President, Global Services Head, AI and Industry Verticals, Infosys
Next Article