Google has unveiled an experimental version of Gemini 2.0 that shows its step-by-step reasoning process. The model performs well on early human-rated benchmarks.
Google has released an experimental version of Gemini 2.0 that shows users how it arrives at its conclusions. According to Google AI chief Jeff Dean, the system has been trained to display its reasoning process as it works through problems.
Built on Gemini 2.0 Flash's architecture, the new model focuses on making its problem-solving steps visible to users. Google AI developer Noam Shazeer says this approach helps the system tackle complex tasks more systematically.
Initial tests by independent benchmarking site lmarena.ai show strong performance, with the model leading in several categories, including maths, creative writing and visual tasks. However, these rankings don't include OpenAI's full o1 model, which should have the upper hand from what we've seen so far. Flash Thinking is likely Google's answer to o1-mini, and there is a more powerful version like Pro or Ultra Thinking.
Google has made Flash Thinking available to developers through the Gemini API in both Google AI Studio and Vertex AI. The company emphasizes that this is an early release and encourages developers to test the model for free and provide feedback.
Google joins industry shift toward AI test-time compute
Like its competitors, Google is moving away from simply making AI models bigger. The company reportedly struggled to achieve meaningful improvements by adding more training data to Gemini 2.0, leading to a new approach that gives models more time to process information during use.
Google's indirect acquisition of Character.ai for up to $2.5 billion brought back renowned AI researcher Noam Shazeer, one of the original authors of the influential Transformer Paper, to focus on what the industry calls "reasoning models."
Flash Thinking appears to be the first result of Shazeer's work. Instead of focusing on pre-training with massive amounts of data, the system dedicates more computing power to what the industry calls "inference time" - when the model is actually solving problems. OpenAI is taking a similar approach with its o1 model.
A recent Hugging Face study supports this new direction, showing that smaller models with better processing capabilities can sometimes match or outperform much larger systems. In one test, a Llama model with just one billion parameters performed as well as a model eight times its size.
