Gemini 3 Pro Preview (high) is a flagship multimodal language model developed by Google, released as part of the Gemini 3 series in late 2025. This specific configuration utilizes the "high" thinking level setting, which enables extended internal reasoning to handle complex analytical tasks, scientific problem-solving, and advanced agentic workflows. It features a context window of 1 million tokens, allowing for the processing of massive datasets, entire code repositories, and long-form video content up to one hour in length.
Built on a Mixture of Experts (MoE) architecture, the model contains over 1 trillion total parameters, with approximately 15 to 20 billion activated during a typical query. This architecture is designed to balance high-level intelligence with inference efficiency, although the high-reasoning mode is notably more verbose and carries a higher token cost than standard configurations.
The model excels in agentic coding and multimodal understanding, demonstrating high performance on benchmarks such as GPQA Diamond, MMLU-Pro, and SWE-Bench. Its capabilities include native support for text, images, audio, and video, with specific optimizations for structured outputs, tool calling, and long-horizon planning.