LLM

Gemini 2.5 Flash Unleashed: Smarter AI with可控的 Thinking Budgets for Developers

Google unveils Gemini 2.5 Flash with enhanced reasoning capabilities, allowing developers to control the 'thinking budget' for smarter and more efficient prompts.

4 months ago

Gemini 2.5 Flash with ‘Thinking Budget’ Rolling Out to Developers

Overview

Google has introduced a preview of Gemini 2.5 Flash, an advanced version of its AI model that includes reasoning capabilities. This new feature, known as the 'thinking budget,' allows developers to control how much reasoning and processing the model performs before generating output. The aim is to balance cost and quality, making it suitable for a wide range of applications.

Key Features

Thinking Budget: Developers can set the number of tokens the model can generate while thinking, ranging from 0 to 24,576 tokens. This parameter can be adjusted via a slider in Google AI Studio and Vertex AI or through an API parameter.
Improved Reasoning: The model can break down complex tasks, plan responses, and improve accuracy, especially for multi-step reasoning tasks like math problems and research analysis.
Cost Control: Setting the thinking budget to zero matches the cost and latency of Gemini 2.0 Flash. If no budget is specified, the model automatically decides based on task complexity.

Specifications

Rate Limits: 1000 RPM / 10,000 RPD (Paid Tier), 10 RPM / 500 RPD (Free Tier)
Knowledge Cutoff: January 2025
Input Modalities: Text, Images, Video, Audio
Output Modalities: Text
Context Window: 1 million tokens
Max Output Length: 64K tokens

Examples of Reasoning Levels

Minimal Reasoning:
- Translation: “Thank you” in Spanish
- Fact-based questions: How many provinces does Canada have?
Medium Reasoning:
- Probability calculations: What’s the probability that two dice add up to 7?
- Scheduling: Creating a basketball schedule around work hours
High Reasoning:
- Detailed analysis and summaries

Availability

Google AI Studio and Vertex AI: Available for preview to developers.
Gemini App: The experimental version of Gemini 2.5 Flash is also coming to the Gemini app, with automatic adjustment based on prompt complexity. End users won’t have manual control over the thinking budget.