Swift 1.5 vs ThinkingCap 3.8 vs Qwen3.8 27B: Accuracy and Token Efficiency

Key points

• That comparison looked at how much accuracy we gain by allowing the model to think more, and how many additional tokens it generates to get there.1
• Swift 1.5 and ThinkingCap 3.8 are two fine-tunes of Qwen3.8 27B designed to reduce unnecessary reasoning.1
• In another article, I found that the ThinkingCap version of Qwen3.6 worked very well, preserving most of the model’s accuracy on non-agentic tasks while significantly reducing the thinking traces.1
• In this article, I compare Swift 1.5, ThinkingCap 3.8, and the original Qwen3.8 27B at Xhigh, looking at accuracy and token efficiency.1

Sources (1)

• [1]Swift 1.5 vs ThinkingCap 3.8 vs Qwen3.8 27B: Accuracy and Token Efficiency
  The Kaitchup (Benjamin Marie) · Oct 13, 04:24 PM

  In a previous article, I compared the reasoning efforts of Qwen3.8 27B, from thinking disabled to Xhigh.

  That comparison looked at how much accuracy we gain by allowing the model to think more, and how many additional tokens it generates to get there.

Extractive summary: sentences quoted from the sources.