ukisai/Swift-Qwen3.8-27b
Swift-Qwen3.8-27B is a 27 billion parameter language model developed by UkisAI, derived from Qwen3.8-27B. This model is optimized for reasoning efficiency, using 58.3% fewer "thinking tokens" while maintaining near-identical performance (less than 1% loss) and achieving a 1.95x speed-up on various tasks. It is particularly suited for applications requiring faster inference and reduced computational overhead in reasoning-intensive scenarios, including general reasoning, mathematics, multimodal tasks, and agentic coding.
Loading preview...
Overview
Swift-Qwen3.8-27B is a 27 billion parameter model developed by UkisAI, built upon the Qwen3.8-27B architecture. Its primary innovation lies in its reasoning efficiency, achieved by fine-tuning to penalize specific "reasoning-marker tokens" that trigger overthinking in the base Qwen model. This approach results in significantly shorter reasoning traces and fewer overthinking errors.
Key Capabilities & Differentiators
- Enhanced Efficiency: Swift-Qwen3.8-27B uses 58.3% fewer thinking tokens compared to its base model, leading to a 1.95x speed-up across various tasks.
- Maintained Performance: Despite the token reduction, the model retains near-identical performance with less than 1% loss on evaluated benchmarks.
- Consistent Savings: Thinking-token savings persist across different
reasoning_effortsettings (xhigh, medium, low) of the Qwen3.8 model. - Quantization Benefits: The model is designed for quantized deployment, retaining token savings in INT4 evaluations and, for AIME 2026, matching or improving accuracy while reducing output-cap failures by 31-33%.
- Broad Application: Benchmarks demonstrate its efficiency across general reasoning (GPQA-Diamond, MMLU-Pro, C-Eval, IFBench), mathematics (AIME 2026, HMMT), multimodal tasks (ERQA), and agentic coding (Terminal-Bench 2.1, LiveCodeBench v6).
When to Use This Model
This model is ideal for use cases where:
- Inference speed is critical for reasoning-heavy tasks.
- Computational resources are constrained, benefiting from reduced token usage.
- Quantized deployment is desired for lower memory footprints.
- Applications involve general reasoning, mathematical problem-solving, multimodal understanding, or agentic code generation where efficiency is paramount.