Qwen3-VL-32B

Qwen3-VL-32B is Alibaba's 32 billion parameter dense vision-language model, released in October 2025. Unlike the mixture-of-experts trend, it uses a dense architecture for predictable deployment, and pairs strong visual perception with text reasoning across images, documents, and long video.

It offers a 256K token context window, OCR in 32 languages, and fine-grained spatial reasoning, with Instruct and Thinking variants for direct answers or extended chain-of-thought. Weights are released under Apache 2.0, making it one of the most capable open vision models at this size.