gemini-2.5-pro

Basic Information

Gemini 2.5 Pro is Google DeepMind’s flagship multimodal large model, a core product in the Gemini 2.X series. Trained on the TPUv5p architecture, it features an ultra-long context window of over 1 million tokens. The model supports multimodal inputs including text, images, audio, and video, and can process video content lasting up to 3 hours. It is now openly accessible via Google AI Studio and Vertex AI, making it well-suited for developer production-level applications and meeting the complex task requirements of professional users.

Key Features

• Top-tier performance across all scenarios: achieving a 74.2% score on LiveCodeBench coding tests, an 88.0% accuracy rate on the AIME 2025 math competition, and an 86.4% reasoning score on GPQA (Diamond Level)—breaking industry records in multiple benchmarks.
• A comprehensive upgrade in multimodal capabilities: the model can convert videos into interactive applications, precisely identify key information in long-form videos, and generate structured outputs such as SVG from images—performance far surpassing previous generations.
• Natively supports tool calls and agent workflows, enabling it to autonomously complete long-term tasks like codebase-level development and solving complex mazes. It can also integrate with Google ecosystem tools to enhance efficiency.

Technical Highlights

• Adopts a Sparse Mixture of Experts (MoE) architecture, dynamically routing tokens to dedicated expert modules to strike a balance between performance and computational efficiency.
• Innovative dynamic reasoning mechanism adjusts the thinking budget based on task complexity, supporting multi-hypothesis generation and self-criticism to reduce hallucination rates.
• Enhanced through adversarial training for robust safety protection; it remains below high-risk thresholds in areas such as cybersecurity and biosecurity risks while minimizing over-rejection and improving practicality.

Note: Gemini native format calls are now supported.