gemini-3-flash-preview

Gemini 3 Flash is a lightweight, flagship AI model launched by Google, primarily positioned as a “high-speed response engine for complex tasks”. It’s designed specifically for high-throughput, low-latency scenarios, and while maintaining powerful multimodal understanding capabilities, it significantly optimizes inference costs and speeds, delivering Pro-level inference performance at one-quarter the cost, thus breaking the impossible triangle of speed, intelligence, and cost. It covers all use cases—from developers and enterprises to ordinary users—enabling ultra-fast responses from high-frequency development to everyday interactions, making it ideal for scenarios such as financial risk control, legal document analysis, and intelligent customer service agents.

• Core breakthrough: No longer needing to compromise among “speed, intelligence, and cost”—it’s 3 times faster than the previous-generation Gemini 2.5 Pro and outperforms it in terms of overall performance; its price is only 1/4 of Gemini 3 Pro, while also supporting cost-optimized solutions like context caching and batch processing.
• Superior coding capabilities: Achieving a 78% score on the SWE-bench Verified coding benchmark, surpassing both the 2.5 series and Gemini 3 Pro, making it the top choice for cost-effective agent coding.
• Enhanced multimodal capabilities: New advanced visual and spatial reasoning abilities have been added, along with code execution functionality (scalable, counting, editing visual inputs). Video analysis speed is 4 times faster than Gemini 2.5 Pro, and OCR accuracy has been significantly improved.
• Cost optimization solutions: Enabling context caching can save up to 90% of repeated token costs; using the Batch API for asynchronous processing further reduces costs by another 50%, while also increasing call rates.
• Large-scale deployment-friendly: Supports an input context of up to 1 million tokens, allowing it to handle large volumes of data such as long documents and video streams while maintaining low latency (in synchronous scenarios, it supports production-level rates).
• Creative and efficiency tools: Supports voice-generated app prototypes (a functional framework can be created in just minutes even with zero coding experience), video content summarization, multilingual Q&A, and generates visual outputs for various needs such as weather checks and travel planning (e.g., schedule tables with calendars).

───────────────────────────────────────────────────────────────────

Core Capabilities

⚡ Ultra-fast response: Delivers high-speed output at 218 tokens per second, ideal for real-time conversations, streaming processing, and high-concurrency API calls, providing “instant results” for tasks such as processing 512-pixel images and generating simple code.

🧠 Flexible inference control: The ‘thinking_level’ parameter supports both “low” (fast thinking) and “high” (deep thinking) modes, precisely balancing speed, cost, and inference depth.

👁️ Comprehensive multimodal processing: Natively supports parsing images, audio, and videos, enabling cross-modal tasks without requiring additional models. Easily handles advanced visual reasoning, video knowledge extraction, high-precision PDF parsing, and complex chart synthesis. 📊 Handling ultra-long contexts: With a 1M-token input window and 64K-token output capability, it can analyze entire financial reports, technical documents, or lengthy novels in one go.

🛠️ Deep integration with toolchains: Built-in Google Search and code execution features automatically verify facts and retrieve the latest data, ensuring that generated content is accurate and reliable.