实测证实:两款微调版 Qwen3.8-27B 推理 token 减约 40% 几乎不掉分
returnity · reddit · 2026-09-25
An independent Aider eval suite comparison of ThinkingCap-Qwen3.8-27B, Swift-Qwen3.8-27B, and vanilla Qwen3.8-27B validates both fine-tunes' claims of 40% reasoning token reduction with minimal performance loss.
Key results (2 runs per model, ±2-3% error bars):
- ThinkingCap scores identically to the original (27.1% first-try / 77.6% retry pass) while median completion tokens drop from 12,547 to 7,436, and wall time nearly halves (777s vs 1,481s per case)
- Swift posts the highest first-try pass rate (30.8%) with 7,301 median tokens
Nuances:
- ThinkingCap uses 8.5% more mean tokens due to a long tail of overthinking cases; Swift reduces more uniformly
- Both spend more tokens on failures than successes, more pronounced for Swift (13.2k vs 5.9k)
- By language: ThinkingCap trails Swift 8% on C++ but leads by 4% (JavaScript) and 6% (Python)
- ThinkingCap is the only model with a perfect well-formed-diff score
「模型」频道最新
- DeepMind 称 Gemini 智能体 4 周完成 Signal 加密协议数学验证 — Polymarket · 2026-11-11
- 爆料：Google 新模型阵容曝光，含 Flash 迭代与视频模型 — koltregaskes · 2026-11-11
- 爆料:OpenAI Astra 6.2 或将临近,已现于 Codex 日志与谷歌云配额 — haider1 · 2026-11-11
- 同一段提示词实测：Fable 5.2 花时 3 倍费用 2 倍，换来导演级表达 — iamfakhrealam · 2026-11-11
- Krafton 开源 1.5B 端侧模型 Raon-1 nano,多项基准超同尺寸对手 — DimitrisPapail · 2026-11-11
- 打错字会让 LLM 不舒服？tokenization 引发有趣讨论 — menhguin · 2026-11-11