Qwen3.5 Slows, Barely Beats Qwen3-VL

Hot take: this looks like an architecture tax, not a free quality upgrade. If the speed hit is real on your stack, Qwen3.5’s marginal accuracy gain is probably not worth it for extraction workloads.

• –Qwen3.5-2B is a vision-capable causal LM with a newer hybrid stack, so extra compute overhead at train and decode time is plausible even at the same size.
• –Qwen3-VL-2B is already a multimodal model, but this report suggests the newer family is not automatically the better throughput choice for OCR-style pipelines.
• –For image-to-JSON, throughput and latency usually matter more than a 1% bump unless that bump materially reduces downstream manual correction.
• –Before concluding it is model-only, verify identical image preprocessing, resolution, prompt format, and decoding settings; those can swing multimodal latency a lot.
• –If your held-out eval backs this up, Qwen3-VL is the pragmatic pick and Qwen3.5 is the “better on paper, worse in production” option.