The narrative around large language models has always been bigger is better. More parameters, more training data, more capability. That story is still true for general-purpose tasks. But for many enterprise use cases, it's increasingly not the right frame. Small, focused, fine-tuned models are delivering better results at dramatically lower cost and latency — and organisations are taking notice.
What "Small" Means Now
The definition has shifted. Models in the 1B–14B parameter range — Phi-4, Gemma 3, Llama 3.2, Mistral Nemo — are capable of tasks that required GPT-4-class models two years ago. Running on a single GPU, with latency measured in milliseconds and cost measured in fractions of a cent per call, these models are economically viable for high-volume production workloads where frontier models are not.
The Fine-Tuning Advantage
A general-purpose frontier model must serve everyone. A fine-tuned small model serves you. When you fine-tune on your domain vocabulary, your output format, your reasoning patterns, and your edge cases, a 7B model will routinely outperform a 70B model on your specific task. We've seen this repeatedly in production: a fine-tuned Llama model for a legal document extraction task outperforming GPT-4 on the client's own evaluation benchmark.
The Cost and Privacy Case
Running a frontier model via API means your data leaves your infrastructure. For regulated industries — finance, healthcare, legal — this is often a non-starter. Small models can be self-hosted on-premises or in a private cloud, keeping data entirely within your boundary. Add in the inference cost difference — often 100x cheaper per token than frontier API calls — and the economics of small models become compelling for any high-volume workflow.
When to Use Each
Frontier models remain the right choice for general reasoning, complex multi-step tasks, and use cases where you don't have enough training data to fine-tune. Small models excel at well-defined, repetitive tasks with consistent inputs: classification, extraction, summarisation, format conversion, and domain-specific generation. The pragmatic approach is to use the smallest model that meets your accuracy bar — and to measure that bar rigorously rather than assuming bigger is better.
