Qwen3.5 Review

Alibaba shipped Qwen3.5 in February 2026 alongside a proprietary Plus variant, and the open one is the interesting half. 397 billion parameters with 17 billion active per token is a sensible size in a year when rivals were publishing trillions, and the architecture is unusual: gated delta networks paired with a sparse mixture of experts, sixty layers, 512 experts with ten routed and one shared, and a hybrid of gated and linear attention. That combination is what lets it hold 262,144 tokens natively and stretch past a million with YaRN scaling without the bill turning into theatre.

The claim Alibaba pushed hardest was agents, specifically that the model can operate desktop and mobile applications rather than merely describe them, which is the direction the whole Chinese chatbot market turned this year. The benchmarks back the general competence: 87.8 on MMLU-Pro, 70.4 on SuperGPQA, 85.0 on MMMU for vision, 76.4 on SWE-bench Verified. None of those top a frontier chart and several sit within touching distance of one.

Apache 2.0 is what makes it worth choosing over the closed alternatives. No revenue threshold, no negotiation, no clause that turns up later. Support for 201 languages and dialects is a real differentiator for anyone working outside English and Chinese, and the vision and video understanding are native rather than adapters. The reservation is that Alibaba's own best model is not this one: Qwen3.8-Max followed in August and stayed proprietary, which is where the ceiling now lives. Weights are free on Hugging Face, and hosted access runs through Alibaba Cloud and Qwen Chat.