Model Overview

INFLogic-Qwen2.5-32B-RL-Preview is a 32.8 billion parameter language model developed by infly, specifically designed to enhance logical reasoning. It is fine-tuned from the DeepSeek-R1-Distill-Qwen-32B base model using a proprietary logical reasoning dataset and a reinforcement learning with verifiable rewards (RLVR) approach.

Key Capabilities & Performance

• Enhanced Logical Reasoning: Achieves state-of-the-art performance among open-source LLMs on the ZebraLogicBench as of March 27th, 2025.
• Benchmark Scores:
  • ZebraLogic: 85.1 (significantly outperforming its base model DeepSeek-R1-Distill-Qwen-32B at 68.7).
  • MATH-500: 95.6
  • GPQA: 65.7
• Fine-tuning Method: Utilizes reinforcement learning with verifiable rewards (RLVR) to improve reasoning abilities.

Good For

• Applications requiring strong logical deduction and problem-solving.
• Tasks involving complex analytical reasoning, such as those found in mathematical or puzzle-based challenges.
• Researchers and developers focused on advancing AI capabilities in logical inference.