Google has introduced Gemini 2.0 Flash Thinking Experimental, an AI reasoning model available in its AI Studio platform. This experimental model is designed to handle multimodal tasks such as programming, math, and physics by reasoning through complex problems and explaining its thought process. It builds upon the Gemini 2.0 Flash model and aligns with similar models, including OpenAI’s o1.
The model uses a structured approach where it breaks down prompts into smaller tasks, analyzes related contexts, and synthesizes the most accurate responses. Despite these capabilities, its reasoning can be inconsistent, as demonstrated by errors in simple tasks like counting letters in a word. It also supports input limits of up to 32,000 tokens, accommodating both text and image inputs, and produces outputs capped at 8,000 tokens in text-only format. The model requires increased inference time computation, leading to slower response times that range from seconds to minutes. It lacks built-in tools for functionalities like search, code execution, or JSON mode, and the accuracy and completeness of its responses may vary. Gemini 2.0 Flash Thinking Experimental also requires longer processing times, a tradeoff for its reasoning abilities.
Jeff Dean, Chief Scientist of Google DeepMind, stated:
The model’s design leverages extended computation during inference to improve reasoning outcomes.
And Logan Kilpatrick, AI Studio’s product lead, described the release as:
An initial step in Google’s exploration of reasoning-focused AI.
The launch follows the recent trend of reasoning models in AI, with competitors such as DeepSeek-R1 and Alibaba’s Qwen also entering the field. These models aim to improve the accuracy and reliability of generative AI systems but come with high computational costs and performance challenges, especially as traditional scaling methods for AI have shown diminishing returns.
Developers can access the model through the Gemini API (v1alpha) or the Google GenAI SDK, which enables integration into various applications, with support for text and image inputs and a focus on transparent reasoning workflows. As a research-oriented release, the model has specific limitations, including token limits and the absence of built-in tool integration.
