By Editorial — IT Researches
The narrative arc of the last three years has been dominated by ever-larger frontier models, and rightly so — the capability gains have been real. The quieter and arguably more important story has been the steady improvement of small models, in the 1- to 7-billion-parameter range, that now solve a non-trivial share of the problems we are asked to solve in 2025.
The cost story is the easy one to tell. A small model running on commodity hardware is somewhere between one and two orders of magnitude cheaper per token than a frontier model called over an API, and the cost gap widens at high volumes. For workloads that are inherently high-volume — content moderation, internal search, classification at scale — the choice has stopped being interesting; the small model wins, and the only question is which one.
The more interesting story is about control. A small model that runs in your tenancy, on your hardware, with your weights pinned, is a system you can reason about. You can audit it; you can fine-tune it without leaking proprietary data to a vendor; you can guarantee its availability without dependence on someone else's rate limits; you can measure its drift over time without negotiating with a third party. For workloads that touch sensitive data or that have to satisfy a regulator's questions about deployment topology, that control is no longer a nice-to-have.
The third story is about specialisation. A small model fine-tuned on a few thousand domain-specific examples will, for a wide class of tasks, beat a frontier model called with a few sentences of system prompt. The gap is widest precisely where the value is highest: enterprise verticals with their own vocabulary, idiosyncratic workflows and documents that would never appear in a public training corpus.
Where small models still struggle, in our experience, is at the long tail of capability — multi-step reasoning that requires holding a great deal of context, generation that has to satisfy stylistic constraints across many pages, the genuinely open-ended dialogue task. Frontier models have a real edge there and probably will continue to. The right architecture in 2025 is to use them where they are needed and to use small models for everything else.
The trap is to treat the choice between small and large as a binary. The systems that we have shipped that have aged best are the ones that route between models — small for the high-volume, narrow tasks; large for the rare, hard ones — with the routing decision itself being a cheap and inspectable layer on top. That architecture is more work than picking one model and committing to it; it is also, in nearly every deployment we have done, the architecture that is still the right architecture two years on.