Overview

taki555/Qwen3-4B-Instruct-2507-Art is a 4 billion parameter instruction-tuned model, a CoT (Chain-of-Thought) efficient version of the Qwen3-4B-Instruct-2507. Developed by Taiqiang Wu, Zenan Xu, Bo Zhou, and Ngai Wong, this model is the result of research detailed in the paper "The Art of Efficient Reasoning: Data, Reward, and Optimization". Its core innovation lies in its ability to generate accurate reasoning trajectories that are significantly shorter than typical CoT outputs, thereby reducing computational costs.

Key Capabilities

• Efficient Chain-of-Thought Reasoning: Optimized to produce concise yet precise reasoning steps.
• Reward Shaping and Reinforcement Learning: Employs a two-stage training paradigm (length adaptation and reasoning refinement) to achieve efficiency.
• Reduced Computational Overhead: Designed to provide the benefits of scaled reasoning with minimal computational expense.

Training Details

The model was trained using the DeepScaleR-Easy dataset.

Good For

• Applications requiring accurate reasoning with strict computational or latency constraints.
• Tasks where concise and direct explanations of thought processes are preferred.
• Research and development into efficient large language model reasoning techniques.