ThinkingCap-Qwen3.8-Flash-Next Reduces Thinking Tokens by 38.5% with Minimal Accuracy Drop

·2026.10.23 02:02

Key point

The post-trained model achieves token reductions of 22% to 51% across six benchmarks while retaining 88.8% average accuracy.

ThinkingCap-Qwen3.8-Flash-Next is a post-trained version of Qwen3.8-Flash-Next designed to reduce token consumption during reasoning.
The model uses 38.5% fewer thinking tokens on average across six benchmarks, with individual reductions ranging from 22% to 51%.
Average accuracy stands at 88.8% compared to the base model's 89.8%, evaluated at xhigh with identical sampling settings on the same hardware.
BF16, FP8, NVFP4, and GGUF weights are publicly available on Hugging Face without access requests.
Benchmarks cover math, coding, long-context, vision, and agentic tasks.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.