Qwen2.5-VL-32B: A Smarter, Lighter Model

Qwen has released Qwen2.5-VL-32B-Instruct, optimized through Reinforcement Learning, under the Apache 2.0 license. This model is based on the existing Qwen2.5-VL series, and while maintaining the 32B parameter scale, it features significantly boosted performance.

The key improvements are as follows:

• Human preference alignment: Response style has been adjusted to provide more detailed and well-formatted answers.
• Mathematical reasoning: The ability to solve complex math problems has been significantly improved.
• Fine-grained image understanding: Accuracy and analytical capability have been strengthened in image parsing, content recognition, and visual logical reasoning.

According to benchmark results, Qwen2.5-VL-32B-Instruct outperforms comparable models such as Mistral-Small-3.1-24B and Gemma-3-27B-IT, and even surpasses the larger Qwen2-VL-72B-Instruct. In particular, it showed an overwhelming advantage in complex multi-step reasoning-focused multimodal tasks such as MMMU, MMMU-Pro, and MathVista.

Additionally, on MM-MT-Bench, which evaluates subjective user experience, it significantly outperformed the previous generation 72B model, and recorded top-tier performance not only in visual capabilities but also in pure text processing capability among models of the same scale.