Small Models Won Again in 2026, and Nobody Wants to Admit It
The flashy launches are trillion-parameter beasts. The work actually gets done by models that fit on a laptop.
Every launch event is about the biggest model. But the quietly useful story of 2026 is the small model: fast, cheap, and good enough for the boring 90 percent of tasks that don't need a frontier brain.
Decision models are the clearest example. AWS open-sourced Strands Decider 2B, a sub-100-millisecond model that just picks the next step in an agent workflow. You don't need a giant reasoning model to route between five options. You need something instant and reliable, and a small model wins that job.
On my own machine, a local model handles draft replies, classification, and summarization with latency a cloud frontier model can't match. The big model is a consultant I call for the hard 10 percent. The small one is the coworker who does the daily grind.
The trap is assuming bigger is better for every step. Teams burn money and latency sending every request to the flagship when a tiny model would do. The 2026 winners are the ones who route by difficulty instead of defaulting to max.
Expect more of this. The interesting engineering isn't 'make the model bigger.' It's 'decide which model should answer.'