The Qwen3.8-27B Variant Built to Stop Overthinking

Swift-1.5-Qwen3.8-27B-GGUF is a quantized version of the Qwen3.8-27B model, designed to reduce excessive reasoning and speed up processing. UkisAI trained this model to lower the number of thinking tokens by 58.5% on GPQA-Diamond, while increasing its score by 0.31 percentage points compared to the original model. The release also claims a 9.18-fold speed improvement on specific tasks.

The model retains 27 billion parameters, but the exact architecture and hardware requirements are unspecified. To run this model, users must use a llama.cpp-compatible runtime like llama-server.

The key benefit of Swift 1.5 is its lower reasoning-token usage, which is advantageous for coding tasks that require extensive reasoning. In LiveCodeBench v6, Swift 1.5 scored 81.71% compared to 76.76% for the original model, and it used 24.5% fewer reasoning tokens. This makes Swift 1.5 a strong candidate for code generation and problem-solving tasks, provided the outputs can be validated using a compiler or test suite.

The benchmark does not guarantee performance across all programming languages or repository-scale tasks.

Swift 1.5 also showed improvements in tool-using and multi-step agent tasks. On Terminal-Bench 2.1, it scored 72.13% compared to 69.21% for the original model, with mean reasoning tokens decreasing from 52,265 to 43,733. This suggests that Swift 1.5 could be beneficial for tasks that involve a chain of reasoning and tool usage. The model was trained with a focus on long-horizon and agentic tasks.

For general reasoning within a token budget, Swift 1.5 outperformed the base model on GPQA-Diamond, scoring 88.59% compared to 88.28%, while reducing mean tokens from 15,014 to 8,717. This indicates that fewer reasoning tokens can lead to better performance in scenarios where token consumption is a concern. The model is available in different GGUF file sizes, ranging from 8.9 GB to 29.0 GB, allowing users to balance memory usage with performance.