UkisAI Releases Swift Model Boosting Qwen 3.8 27B Inference Speed by 1.95x

UkisAI released the Swift-Qwen3.8-27B model, which maximizes inference efficiency based on Qwen 3.8 27B. By identifying and optimizing overthinking patterns, this model reduces thinking tokens by 58% and improves speed by 1.95x. Accuracy loss remains under 1% on most benchmarks.

Optimization Method and Results

Unlike existing methods that forcibly shorten inference length, this approach applies penalties only to 'anxiety-like' inference loops discovered in PTQ and BF16 models. By generating diverse domain traces in an 8xH100 environment to identify common tokens, accuracy was restored through LoRa SFT and On-Policy Distillation.

Key benchmark comparison results:

• GPQA-Diamond: Accuracy 88.4% → 88.3% (tokens reduced by 58%)
• LiveCodeBench v6: Accuracy 76.8% → 81.6% (tokens reduced by 46%; note that this figure is due to baseline truncation and does not represent pure performance improvement)
• Terminal-Bench 2.1: Accuracy 66.7% → 65.8% (tokens reduced by 39%)
• MMLU-Pro: Accuracy 85.5% → 85.0% (tokens reduced by 28%)

Deployment and Limitations

The model is released on Hugging Face, offering various quantized versions such as GGUF and NVFP4, along with an OpenAI-compatible API. However, on the AIME 2026 benchmark, a 4.6% accuracy loss occurred due to a specific token penalty bug, which is scheduled to be fixed in future updates. This method is a complement, not a replacement, for reasoning effort settings, aiming to remove only unnecessary thinking parts while maintaining xhigh-level accuracy.