Intel / article

Qwen3.8 27B addition in words

Source: simonwillison.net
Date: Published Oct 4, 2026

Context

A locally run Qwen3.8 27B model adds large numbers poorly when it answers directly, and far better when it reasons first. Simon Willison tested adding two positive integers and writing the exact sum in English words, using a compressed build of the Qwen model on a DGX Spark.

With reasoning off, across 5,070 cases, the sum was right 23.57% of the time: 97.04% for one- to three-digit operands but 6.44% for ten- to thirteen-digit ones, even though 96.17% of answers followed the word format. So on long numbers the model usually produced well-formed answers that were wrong.

With reasoning on, it got 167 of 169 pairs right. That run used one attempt per pair rather than 30, because reasoning was slow, and Willison expects a second run would differ. The reasoning traces he shares add digits right to left and carry. This is one model and one run, not a comparison across models.

Key takeaways · AI-distilled

Simon Willison reran Colin Frasier's GPT-4o-era experiment on a DGX Spark, having a Codex session test a Q4_K_M- Qwen3.8-27B with 30 attempts per operand-size combination and reasoning disabled.
Without reasoning the model gave the numerically correct sum in words only 23.57% of the time across 5,070 cases, even though 96.17% of its answers followed the required words-only format.
With medium reasoning it was right on 167 of 169 attempts, but Willison ran one sample per combination rather than 30 and expects a second one-shot run would produce different results.
The reasoning traces show the model aligning both numbers digit by digit and adding right to left with explicit carries, working the sum out longhand before writing it in words.

Terms in this piece · Glossary

quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.

Read the source simonwillison.net