Small AI Models for Money Know When to Stay Unsure

Money teams need fast answers that can be checked. They do not need more chat. These are small language models trained to answer typed questions like choices, scores, or yes-no items. Each model gives a probability in one forward pass. Researchers tested two families called K2-Horizon and Qwen3.5/3.8. The sizes ran from 0.8 billion to 32 billion parameters. All of them used one frozen LoRA recipe on a public decision benchmark.

The training used two new finance collections. One teaches rule reading with exact labels. The other uses real outcomes with 19,640 questions about future returns in walk-forward splits. There was also a sealed 2026 test with 1,360 questions. Narrow fine-tuning on just one type cost general decision skill by 9 to 45 percentage points. This gap shrank as models got bigger.

On real outcome questions, every size from 0.8B to 32B landed at the baseline accuracy of 0.37 to 0.38. There was no hidden edge. The models gave flat, well-calibrated probabilities that followed base rates for each time horizon. This is honest uncertainty. The best calibration happened at 2 billion parameters with an error of 0.026. Bigger models from 9 to 32 billion got worse under prior drift, with errors of 0.043 to 0.044.

A mixed corpus fix puts all skills in one adapter at low cost. It matches or beats the general specialist on the public benchmark. It keeps the policy ceiling and near-specialist calibration on real outcomes. The result is one deployable tool with per-domain temperature settings. Band dispatch analysis shows sharp bands where rules set the answer. It also shows flat bands where reality is unclear and the case should go to human review.