Simon Willison tests Qwen3.8-27B arithmetic in words

The experiment involved 30 attempts per number combination with reasoning disabled, followed by a second run with reasoning enabled. In the reasoning-enabled phase, Willison ran only one sample per pair due to the increased time required. The model correctly calculated the sum in words for 167 out of 169 attempts. Willison noted that because these were one-shot runs, a second iteration would likely yield different results. The report includes reasoning traces showing the model aligning digits and adding from right to left, occasionally correcting its own alignment errors.