Swift Series
#12 v Small & Edge modelyUkisai · 3× · naposledy 26. 9. 2026
The Swift Series from UkisAI is a family of reasoning-efficient language models built on Qwen3.8 base models (including 27B and Flash-Next) as well as Prism ML's Ternary Bonsai 2 27B. The models are trained via Post-Training (penalizing so-called "Reasoning-Marker-Token") to reduce pathological overthinking and thereby generate shorter Reasoning chains with nearly constant accuracy. According to the manufacturer, the current version Swift 1.5 (Qwen3.8-27B) uses 58.5% fewer Thinking-Tokens with 0.35% higher accuracy than the base, resulting in up to 1.95x (in other sources up to 9.18x) speed improvement. The weights are available under the "Swift Open License v1.0" with free usage up to 1 million USD annual revenue as well as via a free