gemini-2.5-flash

Gemini 2.5 Flash

Gemini 2.5 Flash is Google's state-of-the-art workhorse model, specifically designed for advanced reasoning, coding, mathematics, and scientific tasks. It is the next iteration in the Gemini 2.0 series of models, representing a suite of highly-capable, natively multimodal, reasoning models.

Notably, Gemini 2.5 Flash is Google's first fully hybrid reasoning model, giving developers the ability to turn a model's "thinking" on or off via the thinkingConfig parameter. It is optimized for low latency and high-volume traffic, delivering the best balance of price, performance, and well-rounded capabilities.

Architecture

Gemini 2.5 models are sparse mixture-of-experts (MoE) transformer-based models with native multimodal support. Sparse MoE models activate a subset of model parameters per input token by learning to dynamically route tokens to a subset of parameters (experts); this allows them to decouple total model capacity from computation and serving cost per token. They were trained using Google's Tensor Processing Units (TPUs).

Context Window & Token Limits

Native Multimodal Capabilities

Gemini 2.5 Flash is natively multimodal, trained concurrently on text, images, audio, video, and code. This allows for seamless cross-modal reasoning without intermediate conversion steps.

Documents & Text

• Process up to 3,000 files per prompt.
• Maximum 1,000 pages per file.
• Ideal for Large Document Analysis: Process a 1M token archive in one shot without retrieval pipelines or chunking errors.

Images (Vision)

• Maximum images per prompt: 3,000.
• Maximum file size (inline/direct upload): 7 MB.
• Maximum file size (Google Cloud Storage): 30 MB.

Video

• Maximum video length (with audio): Approximately 45 minutes.
• Maximum video length (without audio): Approximately 1 hour.
• Maximum number of videos per prompt: 10.

Audio

• Maximum audio length per prompt: Approximately 8.4 hours (up to 1 million tokens).
• Maximum number of audio files per prompt: 1.

Supported Features

• ✓ Hybrid Thinking Mode (Reasoning)
• ✓ Grounding with Google Search
• ✓ Code Execution
• ✓ System Instructions
• ✓ Structured Output (JSON)
• ✓ Function Calling / Tool Use
• ✓ Implicit Context Caching
• ✓ Explicit Context Caching

Usage & API Integration

Load the model using the official Google Gen AI SDK. You can dynamically enable the reasoning (thinking) feature by specifying a thinking budget.

import { GoogleGenAI } from "@google/genai";

const ai = new GoogleGenAI({ apiKey: process.env.GEMINI_API_KEY });

async function runHybridReasoning() {
  const response = await ai.models.generateContent({
    model: "gemini-2.5-flash",
    contents: "Analyze this complex architectural diagram and write a migration plan.",
    config: {
      // Enable thinking for complex tasks natively
      thinkingConfig: {
        thinkingBudget: 1024,
      },
      temperature: 0.4,
    }
  });
  console.log("Model Output:", response.text);
}

runHybridReasoning();

Performance Benchmarks

Gemini 2.5 Flash is heavily optimized for speed and large-scale deployment. Below are the average throughput and latency metrics based on independent evaluations: