Swift Qwen 3.8 27B Hits 100K+ Downloads on Hugging Face
AI lab UkisAI announced that its open-source Swift Qwen 3.8 27B model reached over 100,000 downloads, becoming the top trending fine-tune on Hugging Face. The model employs a specialized optimization technique that penalizes redundant overthinking, cutting token usage by 58.3% and increasing inference speed by nearly 2x without sacrificing accuracy. Reasoning-focused language models often generate excessively long internal chain-of-thought tokens, which drastically increases compute costs and response latency. Demonstrating that reinforcement learning can train smaller models to think more efficiently rather than simply truncating their responses offers a valuable framework for building faster, cheaper open-source AI. UkisAI achieved these efficiency gains by penalizing pathological reasoning loops during RL fine-tuning rather than directly constraining the model to output shorter responses. The team is preparing to release updated checkpoints, including Swift1.5 and a Flash version, featuring bug fixes and benchmark evaluations across coding and long-horizon tasks.
## BACKGROUND
Modern reasoning LLMs generate step-by-step intermediate thoughts before returning a final answer, which improves accuracy on complex problems. However, smaller open-source models often fall into repetitive or unnecessary thinking loops that waste memory and computational resources. Fine-tuning techniques like reinforcement learning (RL) allow developers to align model behavior and prune redundant reasoning patterns.