جوجل ديب مايند تكشف عن جيميني 2.5 برو: قفزة إلى الأمام في الذكاء الاصطناعي والاستدلال والترميز

Google DeepMind has just announced Gemini 2.5 Pro Experimental, their most intelligent AI model to date, representing a significant advancement in AI capabilities, particularly in reasoning and coding. This new model builds upon the Gemini foundation and introduces enhanced features designed to tackle complex problems more effectively. Gemini 2.5 Pro Experimental is now available for experimentation in Google AI Studio and Gemini Advanced, with wider availability on Vertex AI coming soon. This marks a new era of AI that focuses on deeper comprehension and contextual understanding.

Enhanced Reasoning Capabilities

Gemini 2.5 Pro is designed as a “thinking model”, capable of reasoning through problems before responding, leading to higher accuracy and better overall performance. Key highlights include:

• State-of-the-Art Performance: It tops the LMArena leaderboard, indicating a high level of human preference and sophisticated style.
• Leading Benchmarks: Excels in math, science, and coding benchmarks, showcasing its capabilities in complex problem-solving.
• Humanity’s Last Exam: Achieves a remarkable 18.8% score on Humanity’s Last Exam without tool use, demonstrating advanced reasoning on a dataset designed to capture the human frontier of knowledge.
• Thinking Model: This is enabled through improved post-training and by building reasoning capabilities directly into the model.

Advanced Coding Performance

DeepMind has focused intensely on boosting Gemini’s coding abilities. Gemini 2.5 Pro achieves a substantial improvement over Gemini 2.0, making it a valuable tool for developers.

• Visually Compelling Applications: Excels at creating visually compelling web applications and agentic code applications.
• Code Transformation and Editing: Demonstrates proficiency in code transformation and editing tasks.
• SWE-Bench Verified: Achieves a score of 63.8% on SWE-Bench Verified, the industry standard for agentic code evaluations, with a custom agent setup.
• Single-Line Prompt to Video Game: Can generate executable code to create a video game from a single line prompt, showing incredible reasoning abilities.

Building on Gemini’s Strengths

Gemini 2.5 Pro leverages the core strengths of the Gemini model family:

• Native Multimodality: Can understand and process various data types, including text, audio, images, and video.
• Long Context Window: Features a 1 million token context window (with a 2 million token window on the horizon), allowing it to comprehend vast datasets and tackle complex problems involving diverse information sources.
• Accessibility: Available in Google AI Studio and the Gemini app for Gemini Advanced users. Vertex AI integration is planned for the coming weeks.

Availability and Future Plans

Gemini 2.5 Pro Experimental is currently available for developers and Gemini Advanced users to begin experimenting with. Google DeepMind is encouraging users to provide feedback to further improve the model’s capabilities. Pricing for scaled production use will be announced in the coming weeks. The company aims to integrate these “thinking” capabilities into all future models to support more complex tasks and the development of more capable, context-aware agents.

In conclusion, Gemini 2.5 Pro represents a significant leap forward in AI capabilities, with its enhanced reasoning and coding abilities paving the way for more sophisticated applications and solutions. The focus on building thinking capabilities directly into the model signals a shift towards more intelligent and context-aware AI systems. The ongoing feedback loop and planned integration into Vertex AI indicate Google DeepMind’s commitment to continued improvement and wider accessibility for developers and enterprises.

Source: Google DeepMind