Gemini 2.5 Pro

O Google DeepMind

Założona 2010 · Mountain View, USA / London, UK

Google DeepMind is the merged AI research organisation formed in April 2023 by combining Google Brain (founded inside Google in 2011) with DeepMind (founded in London in 2010 by Demis Hassabis, Shane Legg and Mustafa Suleyman, acquired by Google in 2014). Demis Hassabis leads the unit as CEO. DeepMind authored seminal papers including 'Attention Is All You Need' (Google Brain, 2017), AlphaGo (2016), AlphaFold (2018-2021, awarded the 2024 Nobel Prize in Chemistry), AlphaZero, Chinchilla scaling laws and the Gemini Technical Report. The Gemini family launched in December 2023 (Ultra, Pro, Nano), followed by Gemini 1.5 Pro with 1M+ context and the Mixture-of-Experts architecture in early 2024, Gemini 2.0 Flash (December 2024) and Gemini 2.5 Pro (released March 2025, generally available May 2025) which introduced a built-in 'Deep Think' reasoning mode. Google DeepMind also ships Imagen, Veo, Lyria, NotebookLM and powers AI features across Google Search, Workspace and Android.

Architektura

Sparse Mixture-of-Experts Transformer (natively multimodal, reasoning)

Gemini 2.5 Pro is Google DeepMind's flagship reasoning model released as preview in March 2025 and generally available in May 2025. It is a natively multimodal Sparse Mixture-of-Experts (MoE) Transformer that ingests text, image, audio and video tokens through a shared embedding space, building on the architecture introduced in Gemini 1.5 Pro. Pretraining used Google's TPU v5p infrastructure on a multi-trillion-token corpus mixing web text, code, books, scientific papers, image-text pairs, audio and YouTube video frames, with a knowledge cutoff of January 2025. Post-training combined supervised fine-tuning, RLHF and reinforcement learning against verifiable rewards for math and code, plus a new 'Deep Think' training stage that teaches the model to allocate test-time thinking budgets and emit long internal chains-of-thought before the final answer. Deep Think is exposed to developers as a toggle and was the headline capability at launch, pushing Gemini 2.5 Pro to the top of the Chatbot Arena leaderboard for the first time for Google. The model supports a 1,048,576-token context window with native handling of long PDFs, hour-long videos and full code repositories. Tool use, function calling, structured output and grounding with Google Search are first-class. Safety training included red-teaming under Google's Frontier Safety Framework.

Parametry

Undisclosed (sparse MoE, total parameters in the hundreds of billions, active per-token undisclosed)

Kontekst

1 048 576 tokenów

Możliwości

• 1,048,576 token context window (~1M tokens) with native long-video and long-audio support
• Built-in Deep Think reasoning mode with adjustable thinking budget
• Natively multimodal: text, image, audio and video in a single pass
• Top scores on GPQA Diamond, AIME 2025 and Humanity's Last Exam
• Strong coding performance on SWE-bench Verified and LiveCodeBench
• Function calling, JSON schema, parallel tool calls
• Grounding with Google Search built into the API
• Native execution of code via the Code Execution tool
• Cross-lingual reasoning across 100+ languages
• Available via Vertex AI, AI Studio, and the Gemini app
• Best for: scientific reasoning, long-document and long-video analysis, coding, agentic workflows.

Trening i licencja

Pretrained on a multi-trillion-token mixture of web text, code, books, scientific papers, licensed third-party text, audio waveforms, image-text pairs and video frames. Knowledge cutoff is January 2025. Post-training uses supervised fine-tuning, RLHF, RL against verifiable rewards and Deep Think reasoning training.

Licencja: Proprietary commercial license via Google AI Studio, Vertex AI and the Gemini app. Commercial use permitted under Google Cloud terms.

Testy bezpieczeństwa: Evaluated under Google DeepMind's Frontier Safety Framework v2 covering CBRN, cyber, persuasion and ML R&D risks, with internal red teams and external evaluators.