BABA(US:BABA) Changjiang Securities·2025-11-06 09:17Investment Rating - The report maintains a "Positive" investment rating for the industry [8] Core Insights - On November 3, Alibaba released the early preview of its flagship reasoning model Qwen3-Max-Thinking, which is still under training. The model shows excellent performance in mathematical reasoning, agent programming, and common-sense reasoning tasks, but it also has issues with high token consumption and usage costs. The model is not yet open-sourced, and more versions will be released in the future. The report emphasizes the importance of monitoring the subsequent releases and open-sourcing of Qwen3-Max reasoning model, while being optimistic about the commercial progress of Agents. It recommends focusing on shovel stocks and major players with significant positioning advantages [2][5][11] Summary by Sections Event Description - On November 3, Alibaba launched the early preview of the Qwen3-Max-Thinking reasoning model, which is currently in training. This model is a checkpoint in training and is not yet open-sourced on platforms like Hugging Face or MoDa. Users can try it out on Qwen Chat and Alibaba Cloud API, with text-to-text output available for a limited time for free [5] Event Commentary - The Qwen3-Max-Thinking model achieved a 100% accuracy rate in AIME 25 and HMMT25 tests, marking a significant breakthrough in performance. The model integrates a code interpreter and employs parallel testing computation technology, showcasing strong reasoning capabilities. It effectively combines thinking and non-thinking modes, significantly improving reasoning abilities in various domains. The report notes that the model's breakthrough in mathematical testing signifies a qualitative change in the complex logical reasoning capabilities of large models, which may accelerate commercialization [11] - The model has high token consumption, which is somewhat controllable. Users can switch within a token budget range of 1024 to 81920 tokens to directly control the reasoning length. The report highlights the need to pay attention to the team's future technical breakthroughs in cost reduction [11] - The Qwen3 family has released over ten models, forming a complete matrix of trillion-parameter flagship models, open-source multimodal models, and vertical tools. The report continues to be optimistic about the domestic AI industry chain and recommends focusing on shovel stocks and major players with significant positioning advantages [11]