Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2

VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Mar 19, 2026License:apache-2.0Architecture:Transformer0.1K Open Weights Cold

Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2 is a 27 billion parameter Qwen3.5-based language model fine-tuned by Jackrong, optimized for efficient chain-of-thought reasoning. It achieves significant reductions in reasoning length while preserving accuracy, particularly on tasks requiring logical deduction and problem-solving. This model excels in analytical tasks, coding, and mathematics by distilling concise, reusable reasoning patterns from Claude 4.6 Opus.

Loading preview...

Model Overview

Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2 is the second iteration of a 27 billion parameter Qwen3.5 fine-tune, specifically engineered to enhance the efficiency of chain-of-thought reasoning. This version was trained using 14,000 Claude 4.6 Opus-style general reasoning samples, focusing on transferring concise and reusable reasoning patterns.

Key Differentiators & Performance

  • Reasoning Efficiency: Achieves approximately a 24% reduction in chain-of-thought length compared to its base model.
  • Accuracy & Throughput: Demonstrates +31.6% more correct solutions per token, indicating higher efficiency without sacrificing accuracy.
  • Code Performance: Preserves the base model's accuracy on HumanEval (96.91% pass@1), despite being trained primarily on general-domain reasoning data rather than specialized code-heavy supervision. This highlights its robust and transferable reasoning scaffold.
  • Reasoning Scaffold: Incorporates a streamlined reasoning paradigm, such as "Let me analyze this request carefully: 1..2..3...", to reduce redundant cognitive loops and improve inference efficiency.

Training & Data

The model was fine-tuned using Unsloth and LoRA, with response-only training masked on specific assistant tokens. The distillation data primarily consists of high-quality, filtered reasoning datasets, including:

Intended Use Cases

This model is best suited for:

  • Offline analytical tasks
  • Coding and mathematical problem-solving
  • Logic-dependent prompting where transparent AI internal logic is crucial.

Limitations

  • May occasionally hallucinate external facts during reasoning sequences.
  • Due to its training focus, it might underperform the base model on tasks requiring extensive long-context understanding or highly complex multi-step reasoning beyond its optimized scope.