UkisAI Swift GLM5.3 + GSQ-RCO, Swift Ornith 9B / -57.7% Token Usage with <1% accuracy loss on max reasoning

Hi everyone, Jovan from UkisAI here, Today we are expanding our Swift model series with two models requested by the community. We made Swift GLM-5.3-Flash and Swift Ornith 9B more efficient by figuring out which tokens were linked to pathological overthinking and penalizing them without "atta

中文全文翻译正在补齐，可先查看下方英文原文。

Hi everyone, Jovan from UkisAI here, Today we are expanding our Swift model series with two models requested by the community. We made Swift GLM-5.3-Flash and Swift Ornith 9B more efficient by figuring out which tokens were linked to pathological overthinking and penalizing them without "atta