Skip to main content
Back

UkisAI's Swift 1.5 Makes Qwen3.8 Run 9x Faster for Local AI Agents

UkisAI's Swift 1.5 trims Qwen3.8-27B's thinking tokens by 58.5% while matching or beating base accuracy, with up to 9.18x speedups on some tasks.

Model, Full Access
  • UkisAI released Swift 1.5 Qwen3.8-27B GGUF, a reasoning-efficient derivative of Qwen3.8-27B.
  • Uses 58.5% fewer thinking tokens while scoring 0.35% higher than the base model.
  • Up to 9.18x speedup on some tasks; demo game built in 11.4 min vs 104.6 min.
  • Beats base on LiveCodeBench (81.71% vs 76.76%) and Terminal-Bench 2.1 (72.13% vs 69.21%).
  • Slightly worse on AIME 2026 (96.00% vs 98.67%) and HMMT November 2025.
  • Free under Swift Open License below $1M annual revenue; 22 GGUF quant tiers from 8.9 to 29 GB.

UkisAI has released GGUF quantizations of Swift 1.5, a post-trained derivative of Alibaba’s Qwen3.8-27B. The files target local runtimes including llama.cpp, Ollama, LM Studio, and vLLM. Swift addresses a costly behavior in reasoning models: generating thousands of intermediate tokens before producing an answer. Because those tokens are decoded sequentially, they increase latency, compute use, and serving cost.

Across UkisAI’s evaluation suite, the team reports 58.5% fewer thinking tokens and a score 0.35% above the base model. Its largest wall-clock result came from a prompted 3D game demo: Qwen3.8-27B finished in 104.6 minutes, compared with 11.39 minutes for Swift, a 9.18x speed-up. These figures come from UkisAI’s tests, and completion time will vary with hardware, runtime, quantization, context length, and sampling settings.

A targeted cut to overthinking

UkisAI’s post-training pipeline identifies tokens associated with unproductive reasoning loops and applies targeted penalties to them. Reinforcement learning and on-policy distillation, or OPD, then recover task accuracy. A blanket length penalty can teach a model to stop before completing the reasoning it needs; the targeted approach aims to remove repetitive paths while preserving useful intermediate work.

Swift 1.5 extends the pipeline introduced with Swift 1.0. The new training run emphasizes long-horizon coding, tool use, and agent workflows in which token costs accumulate across many model calls.

Coding gains, math regressions

UkisAI evaluated Swift and the base model with a matched five-seed protocol, which applies the same sampling setup and averages variation across random seeds. The reported results favor Swift on coding, terminal-agent, science, and Chinese-language evaluations.