bottlecapai/ThinkingCap-Qwen3.6-27B
ThinkingCap-Qwen3.6-27B is a 27 billion parameter language model developed by BottleCap AI, fine-tuned from Qwen3.6-27B. This model is optimized for token efficiency in reasoning tasks, achieving up to 90% fewer thinking tokens while preserving the original Qwen answer quality. It excels in general reasoning, math, code, and agentic use cases, making it suitable for applications requiring efficient and high-quality thought processes.
Loading preview...
ThinkingCap-Qwen3.6-27B Overview
ThinkingCap-Qwen3.6-27B, developed by BottleCap AI, is a fine-tuned version of the Qwen3.6-27B model. Its primary innovation lies in significantly reducing the number of "thinking tokens" required for complex tasks, achieving an average reduction of 50% and over 90% in best cases, without compromising the base model's answer quality or style.
Key Capabilities & Performance
- Token Efficiency: Demonstrates substantial reductions in thinking tokens across various benchmarks, including GPQA-Diamond (↓ 67.8%), SuperGPQA (↓ 58.4%), MMLU-Pro (↓ 53.7%), and C-Eval (↓ 47.1%).
- Reasoning & Problem Solving: Maintains strong accuracy in knowledge, reasoning, math, and code benchmarks, such as HMMT and LiveCodeBench, while being more token-efficient.
- Guardrail Preservation: The fine-tuning process ensures that the model's safety behaviors remain intact, refusing harmful or jailbreak prompts at rates statistically indistinguishable from the base model, but with fewer thinking tokens.
- Quality Retention: Rigorous evaluation on both in-domain and out-of-domain datasets confirms that the model preserves the original Qwen's answer quality and style.
Ideal Use Cases
- Cost-Sensitive Applications: Excellent for scenarios where reducing inference costs associated with reasoning tokens is critical.
- Reasoning-Intensive Tasks: Suitable for general reasoning, complex problem-solving, and agentic workflows where efficient thought processes are beneficial.
- Maintaining Quality: For users who require the performance of Qwen3.6-27B but with improved token efficiency for reasoning steps.