Industry executives and experts share their predictions for 2026. Read them in this 18th annual VMblog.com series exclusive.
By Parth Patwari, AI & Data Practice Lead, Deloitte
As we look ahead to 2026, organizations are focused on enhancing their AI investments for efficiency and scalability while enabling secure and reliable data architectures. Large language models (LLMs), the engines behind many generative AI tools, have redefined what’s possible, as they’re powerful in processing large data sets. Yet, an alternative to LLMS, small language models (SLMs), are an increasingly compelling option. SLMs are typically defined as models with around 10 billion parameters or fewer, distinguishing them from their larger LLM counterparts�. For organizations determined to do more with less – SLMs can be advantageous, all the while prioritizing data infrastructure integrity.
This piece explores where SLMs are most impactful, outlines the benefits and challenges surrounding their adoption, and highlights how organizations can scale more efficiently and sustainably in the coming year.
The SLM Advantage
Unlike their larger counterparts, SLMs train on smaller, more specialized, curated datasets. This can be ideal for more specialized use cases: such as those requiring greater precision in outputs, heightened data security, and in bandwidth-constrained or cost-sensitive environments. Recent Deloitte & Intel benchmarking found that running SLMs on CPU-based infrastructure can provide approximately 55-57% cost savings compared to GPU-based deployments, with little or no increase in response time�. For organizations running at scale, these savings can be really meaningful. Additionally, SLMs help reduce energy use and processing demands, offering rapid, accurate results for narrowly defined problems.
Where SLMs Shine
In highly sensitive industries, like healthcare and finance, SLMs offer a chance for these organizations to harness AI without compromising data sovereignty. Deploying SLMs on local CPUs within internal networks or VPCs can allow organizations to keep sensitive data in-house, meeting rigorous privacy and compliance standards-without the need for costly, specialized GPUs or cloud-hosted resources�. Consumer-facing businesses can embed SLMs directly in applications and devices, enabling instant, private, and offline AI experiences, boosting user experiences.
For those organizations that find themselves resource-constrained, SLMs can be a practical way to scale AI.
Challenges with SLM Adoption
While SLMs excel in small, targeted tasks, this smaller scope may hinder performance in areas that require a wider scope or knowledge and reasoning. Additionally, with any AI approach, continuous monitoring and improvement are vital to success, and SLMs require meticulous tuning to maintain precision.
Leaders need to set realistic expectations on where SLMs can thrive, and in what areas LLMs may be a good investment. These decisions are all nuanced and need to be rooted in context to make measurable impact.
The Path Forward
As organizations refine their AI strategies, SLMs will play a pivotal role for those prioritizing privacy and agility. Enterprises who adopt a hybrid approach – establishing a broad reach with LLMs and locking into smaller targeted approaches with SLMs – may come out on top in the year ahead.
##
ABOUT THE AUTHOR
Parth Patwari leads Deloitte’s AI and data practice across all industries in the United States. He has extensive experience working with capital markets and payments institutions to architect, design, and implement large-scale systems that support mission-critical customer, finance, risk, regulatory, and compliance functions. Patwari drives efficiency agendas using AI, analytics, and data management capabilities.
