Google DeepMind

Gemini 2.5 Flash uses a dense transformer with native multimodal attention that processes image patches, video frames, and text tokens through a shared attention mechanism rather than separate modality encoders. Their 1M token context window is achieved through a sparse attention pattern that dynamically adjusts context granularity based on content density. Gemini 2.5 Flash's agentic capabilities include built-in tool use for code execution, Google Search integration, and structured output generation. Their training pipeline uses a mixture-of-experts decoder with a massive mixture-of-deep-agents routing system for multi-task specialization.