Bob Browning's blog

The assumption that bigger AI models always deliver better results is being quietly dismantled — and the economics behind this shift are compelling. Now we have Small Language Models (SLM) to distinguish them from Large language Models (LLM). How do we measure the size of a model?

It is measured in parameters. These are the dials and switches that are adjusted during training of the model. GPT-4 has around two trillion parameters. There are viable SLMs with seven billion parameters. These are genuinely small by comparison, yet for specialized tasks it often wins on accuracy, speed, and cost.

Smaller, specialized AI models can outperform their trillion-parameter counterparts in bounded domains. If your AI needs to answer questions about(say) 4,000 fintech products, you simply don’t need a model trained on Shakespeare and quantum physics. A focused 7 billion parameter model — roughly 28GB of storage, runnable on high-end consumer hardware — can deliver superior results at a fraction of the cost.


The infrastructure picture is nuanced. Standard web hosting won’t cut it — you’ll need GPU-equipped dedicated servers or specialist cloud platforms like Replicate or RunPod. Self-hosting only makes financial sense at high volumes or where latency is critical, such as fraud detection or trading systems. But there’s a meaningful upside: complete data sovereignty, which matters enormously for fintech compliance.

The broader direction of travel seems clear. We’re moving toward a hybrid world where lightweight local models handle routine tasks while complex queries route to cloud services. For businesses evaluating AI infrastructure today, the strategic question isn’t “how powerful is this model?” but “is this model matched to my problem?”

Lean, specialized, and domain-trained will increasingly beat large, general, and expensive.