Gemini 3 Flash Preview

Supported Functionality

Item	Specification
Input	Text, Image, Video, Audio, PDF
Output	Text
Context	1,048,576 tokens
Max Output	65,536 tokens
Vision	✓ Supported
Function Calling	✓ Supported

Description

Gemini 3 Flash Preview is Google's natively multimodal reasoning model, released on December 17, 2025. Built on the Gemini 3 Pro reasoning foundation, it delivers advanced reasoning, agentic, and vibe-coding capabilities at Flash-level latency. Its knowledge cutoff is January 2025, and Google has not disclosed its parameter count. As of August 26, 2026, no shutdown date has been announced for the preview endpoint, but Google recommends migrating to the stable gemini-3.6-flash.

It supports a 1,048,576-token input context, a 65,536-token maximum output, four thinking levels, Structured Outputs, and fully multimodal input. Google's launch results report 90.4% on GPQA Diamond, 33.7% on Humanity's Last Exam without tools, 81.2% on MMMU-Pro, and 78% on SWE-bench Verified. Google also states that it is three times faster than Gemini 2.5 Pro and uses 30% fewer tokens on average over typical traffic.

Key Capabilities

• Configurable Advanced Reasoning: Supports minimal, low, medium, and high thinking levels, with high as the default, to balance quality, latency, and cost by task complexity.
• Agentic Coding: Targets high-frequency iterative development, tool use, and interactive building, with an official SWE-bench Verified score of 78%.
• Native Multimodal Understanding: Jointly analyzes text, images, video, audio, and PDFs, with an official MMMU-Pro score of 81.2%.
• Complex Knowledge Reasoning: Scores 90.4% on GPQA Diamond and 33.7% on Humanity's Last Exam without tools for advanced analysis and planning.
• Million-Token Long Context: A 1,048,576-token input window can hold large codebases, long videos, books, and multi-document collections, with up to 65,536 output tokens.
• Agentic Vision: Plans zooming, inspection, and code-based image processing to ground visual answers in more specific evidence.
• Grounding and Combined Tools: Supports Google Search, Google Maps grounding, File Search, URL Context, Code Execution, Computer Use, and custom functions.

Technical Strengths

Feature	Benefit
Pro Reasoning Foundation with Flash Latency	Combines advanced reasoning with fast responses for interactive coding and agent applications.
Four Thinking Levels	minimal through high let one model cover fast question answering and difficult analysis with flexible reasoning depth.
Native Multimodal Architecture	Processes text, images, video, audio, and PDFs in one context, reducing cross-model transformation and orchestration costs.
Agentic Vision and Code Execution	Actively crops, zooms, and processes visual content for more verifiable analysis of charts, interfaces, and detailed images.
Combined Tools and Structured Outputs	Built-in tools and custom functions can cooperate within one workflow, while structured output supports reliable downstream integration.
Multiple Inference Consumption Modes	Context Caching, Batch, Flex, and Priority inference provide deployment choices for throughput, latency, and cost.