今天去问了官方客服，这个似乎是融合模型了，期待官方的模型卡
Thanks for asking
Here is a transcription of this as it is a very important thing to note (I can't 100% verify the accuracy of the transcription)
Speaker 1 (DeepSeek representative): The announcement has already explained it. The calling method remains unchanged.
Speaker 2: OK
Speaker 1 (DeepSeek representative): The model name has never changed.
Speaker 1 (DeepSeek representative): Regardless of which model is connected to the API backend, or what we update it to, the calling name does not change.
Speaker 2: Alright, understood.
Speaker 2: Thanks
Speaker 1 (DeepSeek representative): It's just that the two models have been merged now. deepseek-chat corresponds to the non-thinking mode, and deepseek-reasoner corresponds to the thinking mode.
Speaker 2: This is the part that's quite confusing, and everyone's questions are focused on this now. So based on what you're saying, my understanding is that the 3.1 model has actually been merged, so chat and reasoner are actually the same model. You just use a different model name depending on whether you want to enable the "thinking" mode. It's not like some competitor models where you control this with a thinking/no_thinking parameter. Is my understanding correct?
Speaker 1 (DeepSeek representative): Correct.
Speaker 2: Okay, thanks. This update is actually quite significant. It would be helpful to explain it more clearly in the group chat.
Speaker 1 (DeepSeek representative): Different model names correspond to whether or not the "thinking" mode is enabled.
good
this means there is a way to activate V3.1's reasoning process, and for now only the base model is release, so maybe in a few days they will release the instruct version + model card + official documents on how to use this model
不过看起来即使你调用 deepseek-chat 现在仍然可能在输出结果里输出 <thinking> 标签——这是我昨天遇到的情况。不清楚这是否是伪思考或者是意外激活了推理模式。
EN:
But it seems that even when you call deepseek-chat, it may still output <thinking>
tags in the results—this was the situation I encountered yesterday. It’s unclear whether this is pseudo-reasoning or accidentally activating inference mode.