Performance Benchmarks and Fine-Tuned Variants of Qwen 3.8

The release of the Qwen 3.8 model family has prompted a series of technical evaluations and the emergence of specialized fine-tuned versions aimed at addressing specific reasoning behaviors. While the base model has demonstrated capabilities in complex code refactoring, users and developers are increasingly focusing on optimizing its reasoning efficiency and inference speed.

Reasoning Optimization and Fine-Tuning Results

Recent technical discussions on community platforms like Reddit have highlighted a trend toward reducing “overthinking” or excessive reasoning loops in the Qwen 3.8-27B model. Several fine-tuned variants have emerged to address this:

  • Swift-Qwen3.8-27B: Developed by UkisAI, this model family claims significant reductions in reasoning token usage. According to UkisAI, the “Swift Flash Next” variant achieves a 63.4% reduction in thinking tokens and a 1.8x speed increase compared to the base model.
  • ThinkingCap-Qwen3.8-27B: Released by BottlecapAI, this version also aims to reduce reasoning tokens. Independent testing using the Aider evaluation suite suggests that ThinkingCap and the vanilla 27B model achieve nearly identical performance in terms of “retry pass” rates (77.6%) and “well-formed diffs” (100% for ThinkingCap vs 99.1% for vanilla). However, ThinkingCap reportedly uses approximately 5,000 fewer median completion tokens than the original model, aligning with its claims of a 40% reduction in reasoning overhead.

Comparative programming benchmarks indicate performance trade-offs between these variants. In testing across C++, JavaScript, and Python, ThinkingCap showed a higher pass rate in JavaScript (85.4%) and Python (78.8%) compared to the base model, but underperformed in C++ (61.5% vs 69.2% for the base). Swift-Qwen3.8-27B demonstrated a higher “first-try pass” rate (30.8%) than both the base and ThinkingCap models.

Inference Efficiency and Hardware Implementation

User reports indicate that the Qwen 3.8-27B model can be run on consumer-grade hardware using various quantization methods. One developer reported achieving output speeds of approximately 65 tokens per second on an NVIDIA RTX 5070 SFF using the Q2_0 quantization level. Other reported performance metrics for the Flash Next model include prompt processing speeds of up to 430 tokens per second on a 12GB VRAM setup.

The cost-efficiency of running these models locally is a noted factor for developers seeking to avoid the recurring costs of proprietary APIs. Unverified user reports suggest that local deployment costs for 1 million tokens can be significantly lower than some commercial providers, with input costs cited at approximately 2.4 cents and output at 70 cents per million tokens.

Agentic Capabilities and Limitations

In agentic workflows, such as those utilizing the Pi agent, the Qwen 3.8 model has been used to perform unsupervised complex refactoring. However, some users have noted technical challenges, including indentation errors during code edits and a tendency for the model to enter long reasoning loops, which can impact response times. While some fine-tuned versions like Swift claim to mitigate these “overthinking” loops, the practical application of these models in autonomous agent environments remains a subject of ongoing community testing and refinement.