Google releases updated experimental version of Gemini 2.0 Flash Thinking for testing
![Gemini 2.0 Flash Thinking - a new tool for developers from Google Google introduces Gemini 2.0 Flash Thinking: a new era of computing](/media/post_big/Gemini-2.0-cover.webp)
In December, Google announced the Gemini 2.0 Flash Thinking model as the first for logical reasoning, and now its updated experimental version is available for testing.
Here's What We Know
This model is based on Flash 2.0, which was introduced earlier that month, and allows you to "show your reasoning clearly" (as in AI Studio). This improves performance when solving more complex problems. It complements other models such as gemini-2.0-flash-exp and gemini-exp-1206.
We've also enabled code execution as a tool, so the model can decide to write and execute code during its response. You can enable it in the sidebar in AI Studio!
- Jack Rae (@jack_w_rae) 21 January 2025
Here's a fun example where the model ballparks the solution with a formula, but writes some python code to arrive at... pic.twitter.com/j8wNp8Yn27
Main features of Gemini 2.0 Flash Thinking Experimental (January 2025)
- Contextual window of 1 million tokens (out of 32k): this is convenient for those who want to "plug in a codebase or request a set of articles with more complex reasoning".
- Support for on-site code execution: for improved tool utilisation.
- Higher output token generation.
- Lower frequency of model discrepancies ("reduced probability of contradiction between opinion and answer").
- Compared to Exp 1219, the new version demonstrates "better performance on maths, science and multimodal tests", including 73.3% on the AIME2024 test (Maths) and 74.2% on the GPQA Diamond test (Science).
Next version of our thinking model series + Code execution + 1M token context! The progress on scaling thinking is incredible and will continue to iterate - available on Google AI Studio! More to come https://t.co/OFacvvK8d9
- Sundar Pichai (@sundarpichai) 21 January 2025
DeepMind CEO Demis Hasabis noted that this "represents very rapid progress since our first release in December". It was noted that they have been developing such planning systems for over a decade, starting with programs like AlphaGo, and are pleased to see a powerful combination of these ideas with the most powerful fundamental models.
Gemini 2.0 Flash Thinking Experimental is available for free testing in Google AI Studio and via the API. Sundar Pichai noted that the progress in scaling reasoning is impressive and will continue, and promised more news in the future.
Source: 9to5Google