Gemini

‘Gemini 2.0 Flash Thinking’ is Google’s first reasoning model

Abner Li | Dec 19 2024 - 9:51 am PT

Google today released an experimental “Gemini 2.0 Flash Thinking” model that “explicitly shows its thoughts” to solve complex problems.

As the name suggests, it is built on “2.0 Flash’s speed and performance.” Google says it is “trained to think out loud,” thus “leading to stronger reasoning performance.”

Competing with OpenAI’s o1, Google shared several demos across physics and probability:

Want to see Gemini 2.0 Flash Thinking in action? Check out this demo where the model solves a physics problem and explains its reasoning. pic.twitter.com/Nl0hYj7ZFS
— Jeff Dean (@JeffDean) December 19, 2024

It’s still an early version, but check out how the model handles a challenging puzzle involving both visual and textual clues: (2/3) pic.twitter.com/JltHeK7Fo7
— Logan Kilpatrick (@OfficialLoganK) December 19, 2024

Curious how it works? Check out this demo where the model solves a tricky probability problem. pic.twitter.com/F3kJv4R9Gy
— Noam Shazeer (@NoamShazeer) December 19, 2024

Gemini 2.0 Flash Thinking is available in Google AI Studio (direct link) and Vertex AI today. You can click “Expand to view model thoughts” and see the reasoning occur in real-time before it provides the final answer. This is “just the first step in [Google’s] reasoning journey.”

It has debuted at “#1 across ALL categories” on the Chatbot Arena LLM Leaderboard. Just yesterday, Google launched made 2.0 Experimental Advanced available in the Gemini app, with Gemini-Exp-1206 also at the top of the leaderboard.

The leap from Gemini-2.0-Flash:

Overall: #3 → #1
Overall (Style Control): #4 → #1
Math: #2 → #1
Creative Writing: #2 → #1
Hard Prompts: #1 → #1 (+14 pts)
Vision: #1 → #1 (+16 pts)

It remains to be seen how this will ultimately launch for end users. These reasoning capabilities will presumably be integrated into the main model down the road, with Google’s framing as being part of the Gemini 2.0 family a good indicator of that. At the moment, we already have a task-specific model with “1.5 Pro with Deep Research.”

Updating…