技术规格

Gemini 2.5 Flash Lite Max Thinking (2025-09-25)

Gemini 2.5 Flash Lite Max Thinking is a high-throughput, multimodal reasoning model engineered by Google DeepMind to deliver advanced cognitive capabilities at a significantly reduced computational footprint. As a specialized variant in the Gemini 2.5 family, it integrates a sophisticated 'thinking' mode that allows the model to perform multi-pass reasoning and internal planning before generating a final response. This architectural design enables the system to handle complex logic, such as mathematical problem-solving and multi-step code generation, while maintaining the low-latency profile characteristic of the Flash Lite series.

The model is built upon a sparse Mixture-of-Experts (MoE) architecture, which optimizes resource utilization by routing tokens through specific expert pathways rather than activating the entire parameter set for every request. This structural efficiency is paired with a massive 1-million-token context window, permitting the ingestion of extensive datasets, complete codebases, or long-form video content without the need for complex chunking or retrieval-augmented generation (RAG) strategies. The model natively supports multiple modalities, including text, image, audio, and video, processing these disparate inputs within a unified transformer framework.

From a deployment perspective, the model offers a flexible 'thinking budget' parameter, allowing developers to dynamically scale the amount of reasoning effort based on specific application requirements. This makes it particularly effective for high-volume production environments where a balance between reasoning transparency and cost-efficiency is paramount. Its primary use cases include automated classification at scale, real-time multilingual translation, and the development of agentic workflows that require consistent instruction-following and concise, accurate outputs.