UkisAI Releases Swift Models with Reduced Reasoning Token Usage
UkisAI released Swift GLM-5.3-Flash and Swift Ornith 9B, achieving over 57% reduction in token usage with minimal accuracy loss on complex reasoning tasks. This was accomplished by penalizing tokens linked to pathological overthinking and applying reinforcement learning alongside on-policy distillation. This release significantly improves the inference efficiency and speed of large language models on reasoning tasks by targeting redundant overthinking loops. Such optimizations lower operational costs and hardware requirements for deploying advanced AI models in production. Swift GLM-5.3-Flash achieved a 61.4% reduction in thinking tokens on MMLU-PRO with an average accuracy loss of only 0.46% across multiple benchmarks, while Swift Ornith 9B reduced thinking tokens by 33.5% with zero accuracy loss. Various quantization formats including GGUF (1bit to 8bit), W4A16AWQ, and NVFP4 are also provided for community use.
## BACKGROUND
On-policy distillation is a knowledge distillation technique where a student model generates its own token trajectories, receiving dense supervision or grading from a high-performing teacher model. Pathological overthinking occurs when large language models continue reasoning far beyond what is necessary, wasting computing resources and sometimes introducing errors.