Benchmarking ThinkingCap and Swift Fine-tunes of Qwen 3.8-27B for Coding Efficiency
An independent benchmark evaluated ThinkingCap-Qwen3.8-27B and Swift-Qwen3.8-27B using the Aider suite to test claims of reduced reasoning token overhead. Results confirmed that both fine-tunes cut median completion tokens by nearly 40% compared to baseline Qwen3.8-27B while preserving high coding pass rates. Reasoning-focused LLMs frequently generate unnecessary chain-of-thought tokens, increasing inference latency and compute costs for users. Demonstrating that targeted fine-tuning can curb overthinking without sacrificing code generation accuracy provides a valuable optimization blueprint for the open-source AI community. Evaluated at Q8_0 quantization in llama.cpp, both fine-tunes matched the original model's ~77% retry pass rate while reducing median solve tokens from 19.3K to ~12.1K–12.8K. ThinkingCap uniquely achieved 100% well-formed code diffs and outperformed Swift on JavaScript and Python, whereas Swift achieved a more uniform token reduction and scored higher on C++ tasks.
## BACKGROUND
Modern open-source reasoning models generate internal reasoning steps before outputting final answers to work through complex logic. Benchmarks such as the Aider evaluation suite test how accurately models can inspect existing code bases and produce precise, properly formatted code edits across multiple programming languages without human intervention.