Qwen Open-Sources Qwen3-VL 235B Vision-Language Model, Challenging Google’s Gemini 2.5 Pro

Details

• Qwen has released Qwen3-VL-235B-A22B, its third-generation vision-language foundation model with 235 billion parameters.
• Two distinct weight packs are available: “Instruct” for general applications and “Thinking” optimized for advanced, chain-of-thought reasoning.
• Internal tests show the Instruct model outperforming Google’s Gemini 2.5 Pro in key image-understanding tasks like RefCOCO, TextVQA, and DocVQA.
• The model delivers multilingual optical character recognition and multimodal generation, demonstrated in Chinese, Arabic, and Hindi.
• Weights, training recipes, and evaluation scripts are openly available under a license permitting commercial use with attribution.
• The architecture combines a 213-billion-token text backbone and a 22-billion-parameter vision adapter, supporting images up to 16,000 pixels in resolution.
• This release advances on the earlier Qwen-VL-110B by doubling context length and cutting inference latency in half using sparse attention techniques.
• Checkpoints and comprehensive documentation are published in the repository, with example notebooks ready for use with Hugging Face Transformers.

Impact

This open-source release brings community models closer to commercial leaders like Google, OpenAI, and Anthropic, fueling pressure for transparency in the sector. Startups and enterprises can now build advanced multimodal AI solutions with fewer barriers, potentially speeding adoption for document and product image analysis globally. If Qwen’s benchmarks hold up to scrutiny, the industry may see renewed investment in open vision-language models and increased debate over the ethics of AI training data.