Available in AI Studio, Google’s prototyping platform for AI technologies, the new model is described as particularly adept at "multimodal understanding, reasoning, and coding."
A model card accompanying the release claims it is capable of addressing challenges in fields such as programming, mathematics, and physics, with the ability to "reason over the most complex problems."
However, early testing suggests that while the technology shows promise, it has significant room for improvement.
The Gemini 2.0 Flash Thinking Experimental model is built on Google’s Gemini 2.0 Flash architecture and is positioned as a rival to OpenAI's advanced reasoning models, such as the o1.
What sets reasoning models apart is their ability to effectively fact-check themselves. By pausing to analyse a problem, generating related prompts, and systematically explaining their reasoning, these models aim to deliver more accurate results compared to conventional AI systems.
This process, however, comes at a cost. Unlike most AI models, which produce answers nearly instantaneously, reasoning models like Gemini 2.0 Flash Thinking Experimental take significantly longer—ranging from seconds to minutes—to arrive at a conclusion.
During operation, the model pauses after receiving a prompt, considering related scenarios and questions before delivering a summarized answer. For instance, when tested on a mathematical problem, the model first generated multiple possible solutions, evaluated them for consistency, and then provided a detailed final response.
While this approach has the potential to reduce errors and increase reliability, users of AI Studio have reported mixed results.
"It's clear that Gemini 2.0 Flash Thinking has the potential to solve intricate problems, but its reasoning process can feel slow and, at times, overly convoluted," said one early tester.
The model represents Google's attempt to address a common challenge in AI: maintaining accuracy across complex problem domains while avoiding common pitfalls such as logical inconsistencies or hallucinated information.
However, given the experimental nature of the release, there are notable limitations. Testers have reported inconsistent performance in certain scenarios and occasional lapses in the logical flow of its reasoning process.
Google has acknowledged the model’s experimental status and s that it is actively seeking feedback from researchers and developers to refine the technology.
