Trending Articles
Trending Topics
Trending
Juan Bernabe Moreno is Director of IBM Research Europe in the UK & Ireland. Here, he explains why small language models could deliver value to businesses in 2025
The impact of generative artificial intelligence (gen AI) and large language models (LLMs) has been extraordinary, with businesses across all industries eager to harness the promised productivity and efficiency benefits.
Despite the widespread excitement, many enterprises are still struggling to realise the full potential of gen AI. Often, this is because organisations are now finding that many LLMs are not ideally suited for enterprise use cases.
There are a few elements to this. LLMs are typically designed as general-purpose models capable of handling a broad spectrum of tasks. However, this “one-size-fits-all” approach often lacks value for enterprises adopting gen AI to address their unique use cases. Many proprietary LLMs also do not offer adequate transparency into their training data, making it riskier for enterprises to add their own data to the model – which is where the real value lies.
Without being able to harness their proprietary data and fine-tune models in a transparent and secure environment, enterprises will get limited business impact from gen AI, deterring further investment and adoption.
For some companies, the rise of smaller language models (SLMs) is already helping to address this issue and we anticipate this will be a major trend in AI adoption in 2025. While smaller foundation models have not garnered the same attention as their larger counterparts, they are a mighty force that can help more businesses realise the value of gen AI.
SLMs do exactly what they say on the tin. They are foundation models trained on smaller data sets with fewer parameters (the variables the models learn in training). While there is no strict definition, SLMs usually have around 2-7 billion parameters and can even be in the hundreds of thousands. LLMs, on the other hand, can venture into the trillions.
While scale is important to unlock a new frontier of insights and innovation, most enterprise AI use cases do not need this vast amount of largely irrelevant data. They need efficient models built on transparent, screened data that can be safely trained on their proprietary data to give accurate, trustworthy results.
SLMs are ideal for supporting specialised tasks, using smaller, focused data sets. This approach provides better data transparency and visibility while minimising privacy and security risks, such as breaches, IP loss and bias. More energy-efficient and cost-effective to develop and maintain, SLMs offer a more accessible starting point for any company looking to leverage gen AI.
Another key advantage of SLMs is their adaptability to smaller devices. Their compact size allows them to be integrated into smartphones or Internet of Things (IoT) devices. With fewer parameters to process and low latency (the ability to respond with minimal delay), they can generate responses faster, making them well-suited for applications that require real-time responses, such as voice assistants or chatbots.
This local processing capability also makes them suitable for scenarios with unreliable or limited internet connections, such as in the agriculture or offshore energy sectors.
Enterprises looking to capitalise on gen AI in 2025 must start by identifying the use cases that could deliver the most value. Then explore which models on the market best fit their requirements, including data quality and budget. By balancing cost efficiencies and transparency with performance and accuracy, SLMs are the perfect entry point for enterprises seeking to accelerate gen AI adoption.
And when it comes to the future development of this technology, we believe the open-source nature of these models is crucial. Not only does it provide greater flexibility for enterprises, but it contributes to a transparent and responsible AI ecosystem, promoting the creativity, collaboration, and competition needed to foster continued innovation.
