Google makes "reasoning model" based on Gemini available

Google's new AI model from the Gemini family can think. At least that's how AI providers currently interpret the term "reasoning". The thinking mode is integrated into Gemini 2.0 Flash, and the model is called "Gemini 2.0 Flash Thinking". However, it is initially only available as an experiment. People with access to Google AI Studio or a direct API connection can test the thinking mode.

Thinking or reasoning in AI models currently means that they go through a kind of answer-finding process. Answers are played back to the model, which checks them, before they are output. This can happen several times. This should make the answers noticeably better and, above all, more correct. Comparing several answers can help the model to detect hallucinations and thus avoid them.

Gemini even shows users the paths of the thought process on request. They are offered as thoughts in the answer window, but with the note that they are experimental thoughts. There are some restrictions for the thinking mode compared to the previously released Gemini versions. The input limit is 32,000 tokens, the output limit is 8000 tokens. Although you can enter text and images, the output is limited to text. Other services cannot be linked to the thinking mode –, such as search or code execution, writes Google in the blog post.

Noam Shazeer apparently played a major role in the new model version, posting several contributions on X. The former Google employee was also jointly responsible for the highly regarded Transformer paper, which sets out the principles of today's large language models. Shazeer had set up his own business with character ai, but the start-up then entered into a deal with Google that brought the leading researchers back to Google and provided for a license model for the models.

(emw)