Qwen3.8 27B addition in words

Blog · Watch/Read · (gathered 2026-10-05)

• Simon Willison’s experiment demonstrates that enabling “reasoning” dramatically improves the accuracy of local LLMs, specifically Qwen3.8-27B-Q4_K_M.gguf, in performing complex arithmetic like adding large numbers and returning the sum in words.
• Running on a DGX Spark, Qwen3.8-27B-Q4_K_M.gguf performed poorly with reasoning disabled, but achieved 167 out of 169 correct answers when prompted to show its reasoning, which included explicit step-by-step calculations and self-correction (“Wait, let me redo this more carefully.”).
• This highlights a crucial practical takeaway: for tasks requiring precision or multi-step computation, instructing an LLM to output its reasoning process can be a highly effective prompting strategy to significantly boost performance, even for models running on local hardware.