Cut Qwen3.8-27B Reasoning Tokens by 40% -- 3.8 'ThinkingCap' benchmarked!

I doubt I'm in the minority here when I say I love Qwen models, but the overthinking is a major timekiller. It was bad in 3.6-27B, and it's worse in 3.8. I know there are some who say, "well that's how it achieves such a good performance/size ratio"... But now there's some definitive proof that's not the case. Introducing Swift-Qwen3.8-27B! This model was clearly inspired by Qwen3.6-27B ThinkingCap , which was the version of 3.6-27B I used as a daily driver before switching to the 3.8 series. For those of you who haven't heard of it, ThinkingCap is a fine-tuned version of 27B that uses about 40% less tokens to accomplish comparable benchmarks and general performance as the original model. It's one of those fine-tunes that actually works. I used it daily for months without any issues, and it saved me countless hours. I had been waiting and hoping that they would release a similar version of 3.8, because it is so slow, despite its impressive performance, but so far none has been forthcoming. However, it looks like UkisAI also enjoyed that model, and took it upon themselves to deliver a sequel. They identified "reasoning-marker tokens that ... trigger overthinking in Qwen’s reasoning rollouts" and penalized them using RL, resulting in fewer overthinking errors. They also employed "a transfer component derived from BottleCap AI's ThinkingCap-Qwen3.6-27B ". The end result is an average of 30-50% fewer reasonign tokens for the same quality outputs on a number of benchmarks (see the model card for all of them). This claim is quite impressive, and I have independently verified their claims and the quality of the model in my own use cases and in coding benchmarks using Aider as an eval suite: Metric Swift-Qwen3.8-27B Qwen3.8-27B Pass1 (%) 30.8 27.1 Pass2 (%) 75.7 77.6 Well-formed diff (%) 98.1 99.1 Completion tokens 7,301 12,547 Seconds/case 750 1,481 Total tokens/solve 12.1k 19.3k As you can see, their claims hold true -- Swift accomplished an equivalent success rate in approximately half the time, using 63% of the tokens! This is a huge win for 3.8-27B users, because of course decode drops off more and more the longer the response gets, which is why the time is halved even though the tokens are closer to two-thirds of 3.8-27B. Anyways, my posts tend to get excessively long so I'll cut it off here, I was just really excited after finishing my eval suite on this model and wanted to share.