ukisai/Swift-1.5-Qwen3.8-27b

UkisAI developed this derivative to reduce reasoning costs while maintaining accuracy. The model uses 58.5% fewer thinking tokens than the base Qwen3.8-27B. It scores 0.35% higher than the foundation model on aggregate benchmarks. According to the publisher, this results in a 1.95x speed-up on specific tasks. The card positions this model as a direct upgrade from Swift 1.0. It targets coding and agentic workloads where efficiency is critical. Users seeking faster inference without significant accuracy loss can adopt this parameter set. The training data is available on Hugging Face for inspection, though it is not used directly out of the box. Independent verification of the performance claims is absent from the source material. The demo shows the derivative completing a game building task in 11.39 minutes compared to 104.6 minutes for the base. Terminal-Bench scores may appear low initially because the model avoids overthinking loops, which alters token usage patterns. The reported 350k+ downloads for the previous version are unverified figures provided by the authors.