PCL-Reasoner-V1
Model Overview
We release PCL-Reasoner-V1, a model trained based on Qwen2.5-32B-Base and undergoes high-performance supervised fine-tuning based on the MindSpore framework and Ascend hardware. After fine-tuning, the model demonstrates significant improvements in mathematical reasoning capabilities. PCL-Reasoner-V1 achieves 85.7% and 84.2% respectively on AIME 24 and AIME 25, which position PCL-Reasoner-V1 among the top-tier models in the 32B parameter class on AIME24/25.
We have fully open-sourced the model weights, dataset and training code. Follow the tutorial below to deploy and explore post-training!
Code
https://github.com/PCL-Reasoner/V1
https://openi.pcl.ac.cn/PCL-Reasoner/V1
Evaluation
We used the Avg@32 metric (averaging 32 sampling attempts per query) for evaluation.
| Parameter Size | Model Name | AIME 24 | AIME 25 |
|---|---|---|---|
| >100B | |||
| DeepSeek-R1 | 79.8 | 70 | |
| DeepSeek-R1-0528 | 91.4 | 87.5 | |
| Qwen3-235B-A22B | 85.7 | 81.5 | |
| OpenAI-o3 | 91.6 | 88.9 | |
| Gemini-2.5-Pro-0506 | 90.8 | 83 | |
| 32B | |||
| Qwen3-32B | 81.4 | 72.9 | |
| QwQ-32B | 79.5 | 69.5 | |
| DeepSeek-R1-Distill-Qwen-32B | 72.6 | 49.6 | |
| Skywork-OR1-32B | 82.2 | 73.3 | |
| AM-Thinking-v1 | 85.3 | 74.4 | |
| PCL-Reasoner-v1 | 85.7 | 84.2 |