UkisAI has released the Swift series of efficient reasoning LLMs, including Swift1.5, based on Qwen and trained to penalize pathological overthinking via RL‑GSPO and OPD, yielding a 63.4% reduction in thinking tokens and a 1.95× speedup while preserving high accuracy. The release also provides the full model family and the requested GSQ‑RCO quantizations for the 27B and Flash‑Next variants, following over 350 k downloads of the initial Swift Qwen 3.8 27B in just 13 days.