Qwen just published the vision language benchmarks of qwen3.5 medium and I have compared Qwen3.5-35b-a3b with Qwen3-VL-235b-a22b, They actually perform close to each other which is insane!

Alibaba's Qwen research team has released surprising benchmark results showing their latest text-only language model, Qwen3.5-35B-A3B, performs remarkably close to their specialized vision-language model, Qwen3-VL-235B-A22B, on multimodal evaluation tasks. The comparison, which went viral in AI communities, reveals that the 35-billion-parameter text model achieves scores nearly matching the 235-billion-parameter vision-language model on challenging benchmarks including MMMU (Massive Multidisciplinary Multimodal Understanding) and MathVista. This unexpected performance parity suggests that pure language models may be developing emergent multimodal capabilities through extensive training, potentially challenging the conventional wisdom that dedicated vision encoders are essential for strong visual reasoning.

Technical analysis indicates the Qwen3.5-35B model achieves this performance through enhanced reasoning capabilities and improved instruction following, despite lacking explicit visual processing components. The implications are significant for AI deployment: smaller, more efficient text models could potentially handle multimodal tasks previously requiring massive, specialized architectures. This development could accelerate multimodal AI adoption by reducing computational requirements and deployment costs. However, questions remain about whether this represents true multimodal understanding or sophisticated pattern matching from text descriptions of visual content in training data.