Attention follows the largest models, which is reasonable — they define what is possible. But the models that have changed what is practical over the past couple of years are the small ones, and they get a fraction of the coverage.
What "small" now means
Models in the range that fits comfortably on a consumer graphics card, or in quantised form on a recent phone, now handle summarisation, classification, extraction, translation and structured output well enough for production use. That was not true not long ago. The gains came from better training data and better training technique rather than from scale, which is why they arrived quietly.
Why it matters more than benchmark position
A model you can run yourself changes four things at once. Data never leaves your infrastructure, which resolves most privacy and residency questions before they are asked. Cost becomes fixed rather than per-request. Latency stops depending on somebody else's network. And the model cannot be deprecated, reprice, or change behaviour underneath you between one week and the next.
For anyone building where connectivity is unreliable or where regulation restricts sending personal data abroad, that combination is worth more than a few points of capability.
The honest trade-off
Small models are worse at long multi-step reasoning, at unusual tasks they were not tuned for, and at anything requiring broad world knowledge. The productive pattern is not choosing one but routing: a small local model handles the large volume of routine work, and the difficult minority is escalated to something larger. That is an engineering decision, and it is where most of the practical value currently sits.
