Jackrong/Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled-v2
Jackrong/Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled-v2 is a 4.5 billion parameter Qwen3.5-4B fine-tune, optimized for efficient general reasoning and structural problem-solving. This v2 iteration focuses on reducing reasoning length while maintaining strong cross-task generalization, particularly from logic and math to programming. It excels in analytical tasks, coding, and logic-dependent prompting where transparent internal logic is valued. The model is designed for resource-constrained environments and agentic workflows requiring economical reasoning.
Loading preview...
Jackrong/Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled-v2 Overview
This model is the second iteration of a 4.5 billion parameter Qwen3.5-4B fine-tune, specifically engineered for efficient general reasoning. It was trained on over 14,000 premium Claude 4.6 Opus-style reasoning samples, emphasizing concise and reusable reasoning patterns. The primary goal of v2 is to enable the model to "think smarter, not longer," significantly improving its reasoning economy and structural efficiency.
Key Differentiators & Capabilities
- Optimized Reasoning Economy: Achieves substantial reductions in average "think length" (33.77% shorter) and improves passes per 10k think characters (over 40% increase) compared to the official Qwen3.5-4B baseline.
- Cross-Task Generalization: Despite general-domain reasoning training, it demonstrates robust transferability to specialized tasks like programming, as evidenced by competitive HumanEval performance.
- Streamlined Reasoning Scaffold: Incorporates a refined reasoning process, such as "Let me analyze this request carefully: 1..2..3...", to eliminate redundant internal loops and reduce verbose over-analysis.
- Distilled from Claude 4.6 Opus: Leverages high-quality distillation data from Claude 4.6 Opus to impart advanced reasoning trajectories.
Ideal Use Cases
- Resource-Constrained Deployment: Excellent for consumer GPUs or local setups where shorter, cleaner reasoning traces reduce latency and memory pressure.
- Agentic Workflows: Improves end-to-end agent speed and lowers cumulative inference costs by providing economical reasoning for multi-step subtasks.
- Open-Source Tool Use & Agent Stacks: Highly practical for lightweight open reasoning systems and autonomous agent projects where reasoning efficiency is critical.
- Analytical & Logic-Dependent Tasks: Best suited for offline analytical tasks, coding, math, and complex logic problems where transparent internal logic is beneficial.