UkisAI Releases Swift-Qwen3.8-27B: 2x Faster Reasoning with Minimal Accuracy Loss
UkisAI has open-sourced Swift-Qwen3.8-27B, a post-trained reasoning model that cuts thinking tokens by 58.3% and speeds up output generation by nearly 2x while maintaining less than 1% accuracy degradation. The team accomplished this by penalizing token patterns associated with redundant reasoning loops and restoring performance using On-Policy Distillation. Reasoning LLMs frequently waste compute resources and generation time on redundant "overthinking" loops that do not improve final output quality. This technique proves that post-training can optimize reasoning length efficiently without forced token truncation, enabling faster and cheaper inference for open-source AI deployment. The release includes HuggingFace weights, GGUF quantizations (Q1–Q8), and community-contributed NVFP4 and W4A16 versions, along with a free rate-limited research API hosted on Nvidia GPUs. Unlike standard context caps or chat template rules, this method specifically penalizes overthinking token patterns observed across both low-precision quantizations and full BF16 models.
## BACKGROUND
Modern reasoning models use chain-of-thought (CoT) prompting to generate intermediate "thinking" tokens before producing a response, but they can easily fall into repetitive calculation loops. On-Policy Distillation is an optimization framework where a student model samples trajectories during training and is guided by a teacher model to refine its output behavior.