Alibaba announces the release of the new Qwen 3 open-licensed AI model family. The flagship model is Qwen3-235B-A22B, positioned as a competitor to DeepSeek-R1, o1, o3-mini, Grok-3, and Gemini-2.5-Pro. The models implement a hybrid approach with two modes: “thinking” – for step-by-step consideration of complex tasks, and “non-thinking” – for fast responses to simple queries. Some variants are built on a mixed experts (MoE) architecture, which increases computational efficiency. Most Qwen 3 models are already available or will be uploaded to Hugging Face and GitHub platforms. The parameter volume of the models ranges from 600 million to 23.5 billion. They can be used via the web version or the Qwen app, available in Russia but requiring a foreign IP address.