It’s fascinating how quickly perspectives shift in AI. Just two weeks ago, an article celebrated small language models (SLMs) as the future for enterprise AI, citing their efficiency, cost savings, and value for business. Now, a new piece argues that advances in large language models (LLMs) have made SLMs nearly obsolete.
The truth is, each model type has unique strengths that suit different needs.
SLMs, with fewer than 10 billion parameters, are well-suited for businesses focused on specific solutions. They’re efficient, quick to deploy, and consume far less energy than larger models—perfect for well-defined industry applications where privacy, simplicity, and cost are top priorities. For many tasks, SLMs are an ideal fit, handling the job with agility at a fraction of the cost.
At the same time, recent techniques, like Google DeepMind’s Relaxed Recursive Transformers (RRTs), are paving the way for more efficient large models. RRTs allow large models to operate with reduced memory and energy, reusing certain layers for a leaner, faster approach. This makes it possible to bring LLM-level sophistication into applications that were once cost-prohibitive, potentially opening new options for businesses that need both depth and efficiency.
For many, using a mix of models will yield the best results. SLMs can handle predictable tasks and routine business processes where speed and cost are primary. Meanwhile, the latest advancements in LLMs provide versatility and insight in more complex scenarios. The key is choosing the right model for each purpose, aligning business needs with each model’s strengths.
These rapid advancements remind us that AI is a moving target. As models evolve, so should our approach to deploying them, staying focused on the best fit between business goals and model capabilities.
