March 6, 2011 - Last Tuesday, a netizen discovered that theByteDanceIts AI assistantBean curdBeing tested on a small scalereflect in depthmodel, but the access is not to the DeepSeek model.
On March 5, 2025, Doubao announced the launch of its "Deep Thinking" reasoning mode, which comprehensively covers Q&A, search, writing and reading scenarios. To enter the "Deep Thinking" reasoning mode, users only need to turn on the "Deep Thinking Beta" switch in the input box.
During last week's test, the person in charge of Beanbag responded that the current test is a different experimental version of its own Deep Thinking model. There is also news that the deep thinking model that Beanbag is testing is based on the Beanbag 1.5 base model.
1AI query found that previously in mid-January, Doubao big model team announced the existence of deep reasoning model Doubao-1.5-pro-AS1-Preview when releasing Doubao 1.5 Pro, and said that it "developed the Doubao deep thinking model under the condition of not using the data of other models at all, and gave full play to the arithmetic advantage of Test Time Scaling through the breakthrough of RL algorithm and engineering optimization". The Doubao Deep Reasoning Model was developed by taking full advantage of the arithmetic power of Test Time Scaling and completing RL Scaling without using any other model data.
Simply put, Deep Thinking mode allows users to visualize the thinking process of the model, which involves the display of the chain of thought, also known as COT (Chain of Thought) - Chain of Thought is simulated, and through training the model outputs the intermediate steps, such as self-questioning and reflection, which is just a sequence of words, but looks like a human thinking process. that look like human thought processes.
In Deep Thinking mode, users can not only see the final answer of the AI, but also observe the complete logic chain of the model to solve the problem, including self-questioning, hypothesis verification, error correction and other steps. For example, when solving math problems, the model will show its entire process from problem disassembly, multi-method verification to the final conclusion.
