Qwen 27B vs Swift: 50% Faster Speeds Come at an Accuracy Cost
TL;DR
UkisAI's Swift is a reasoning-efficient derivative of Alibaba's Qwen3.8-27B that cuts unnecessary thinking tokens. According to UkisAI, mean token usage drops by 24 to 51 percent, with speeds approaching twice the base model. Accuracy holds on GPQA-Diamond and C-Eval, slips on hard math such as AIME 2026 (94.00 vs. 98.67 percent) and improves on LiveCodeBench (81.55 vs.
Nauti's Take
The progress is tangible: cutting thinking tokens by up to half lowers cost and latency, and Swift even scores higher on coding benchmarks. The catch sits in hard math, where accuracy drops, and in a license that kicks in above 1 million dollars in revenue.
Teams running Qwen locally for code tasks should test it; anyone who depends on Apache 2.0 or precise math is better off with the original.
Video
Sources
