A recent evaluation compared three leading models—Gemma 4, Qwen 3.5, and Ministral 3—on their ability to perform vision-related tasks. The test involved analyzing complex visual content, including multi-element scenes and detailed imagery. Among the three, only Qwen 3.5 consistently demonstrated the ability to interpret and respond to visual inputs accurately.

The assessment was conducted in a controlled environment to ensure fair comparisons. Each model was given the same set of visual prompts, with performance measured based on accuracy and relevance of responses. The test revealed that while two models struggled with visual reasoning, Qwen 3.5 handled the tasks with greater precision and fewer errors.

The results showed that Qwen 3.5 achieved a 90 percent success rate in visual comprehension tasks, significantly outperforming the other models. This highlights a clear gap in vision capabilities between Qwen 3.5 and its competitors. The findings suggest that while vision processing is an emerging area for many models, Qwen 3.5 currently leads in this domain.

The performance discrepancy raises questions about the development priorities of different model creators. While some models excel in traditional language tasks, vision capabilities remain uneven. This could influence adoption patterns, as users may favor models with stronger visual reasoning for applications requiring image analysis.

The results indicate that vision processing is still an evolving capability for large language models. While Qwen 3.5 currently leads, the gap suggests opportunities for improvement across the board. As vision tasks become more common in practical applications, model developers may need to focus more on refining these abilities to meet user demands.