Qwen 3.8 27B tests how LLMs handle arithmetic word outputs. The experiment benchmarks addition accuracy across varying digit lengths when models must return the sum spelled out in words rather than digits. Developers can view detailed heatmap charts and raw logs to analyze failure modes in tokenization and numerical reasoning.
Opening Kapyn…