Qwen, the AI research team backed by Alibaba, just released Qwen3.6-Plus, and it's a departure from what made them famous. Unlike their previous models that anyone could download and run, this one is closed-weight, hosted only on Alibaba Cloud's ModelStudio and OpenRouter. The model targets real-world agent applications with a 1 million token context window and scored 78.8 on SWE-bench Verified for coding tasks.
The benchmark comparisons have rubbed some people the wrong way. Qwen compared the model against Claude Opus 4.5 and Gemini Pro 3.0, which aren't the current versions of those products. Critics called it misleading. Defenders say established baselines matter more than chasing the newest numbers. Maybe so, but the optics are awkward for a team that built its reputation on openness.
Technically, Qwen3.6-Plus uses a hybrid architecture combining linear attention with sparse mixture-of-experts routing. That's how it handles those massive context windows without the inference costs ballooning. You can test it free on OpenRouter right now, or pay for enterprise-grade access through Alibaba Cloud.
The shift to closed weights isn't surprising if you follow the economics. Running a million-token model with MoE architecture costs real money. When Qwen released open weights, users could spin up the model on any cloud they wanted. Now Alibaba captures that inference revenue directly. The open-source community feels burned, but for B2B customers who need compliance and security guarantees, this model was never really for them anyway.