Qwen2.5-VL-32B: Smarter and Lighter

GitHub Pages · 1 year ago · 26

Alibaba released Qwen2.5-VL-32B-Instruct, a vision-language model optimized through reinforcement learning with a 32 billion parameter scale under Apache 2.0 license. The model surpasses larger competitors including Qwen2-VL-72B-Instruct on benchmarks like MMMU and MathVista while matching or exceeding performance of Mistral-Small-3.1-24B and Gemma-3-27B-IT. The improved model enables more detailed responses, enhanced mathematical reasoning, and better fine-grained image understanding suitable for complex multi-step visual reasoning tasks.