Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

By Gheorghe Comanici, E. Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit S. Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen and 3,271 more. arXiv.org, 2025

3,609 citations recorded by the source corpus.

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

View Paper: https://www.semanticscholar.org/paper/39d9c3f1cd4bd5069713e50dc7301570575fc055

Citations in the corpus

• Qwen3-VL Technical Report (2025, unknown): “5-Pro’s (Comanici et al., 2025) Thinking mode delivers the best overall performance, but Qwen3-VL-235B-A22B-Thinking is not far behind.”
• Qwen3-Omni Technical Report (2025, unknown): “…al., 2023a; Yang et al., 2024; 2025a; Touvron et al., 2023; Dubey et al., 2024; Li et al., 2023; Liu et al., 2023; Zhu et al., 2023; Bai et al., 2023b; 2025; Chu et al., 2023; 2024), natively multimodal systems have drawn substantial attention (OpenAI, 2024; Comanici et al., 2025; Xu et al., 2025).”
• LVBench: An Extreme Long Video Understanding Benchmark (2024, unknown): “5-Pro [7].”
• Do Agents Dream of False Memories? Black-box Visual Attacks on Long-term Memory in Multimodal AI Agents (2026, unknown): “RAG partially addresses this by externalizing knowledge into a vector index [46, 44, 9, 28, 56], and scaling context windows offers a complementary but incomplete solution [19, 53, 27, 12, 48].”
• Invoice Haystack: Benchmarking Document Retrieval and Visual Question Answering Under Strong Visual Homogeneity (2026, unknown): “Gemini 2.5 Flash [17] analyzes each anonymized image to generate one question-answer pair per document, yielding 2,000 raw candidate pairs.”
• TrajPrism: A Multi-Task Benchmark for Language-Grounded Urban Trajectory Understanding (2026, unknown): “1 9 and Gemini 2.5 Flash [6] to validate scoring consistency, and 100 randomly sampled Porto trajectories are assessed by human annotators to establish a quality baseline.”
• Black-box model classification under the discriminative factorization (2026, unknown): “As model providers increasingly restrict access to only API endpoints [24, 4, 13], black-box interactions are often the only available mode for users and researchers.”
• A Large-Language-Model Framework for Automated Humanitarian Situation Reporting (2025, unknown): “To ensure a fair comparison, we standardize the underlying model architecture across systems, employing Gemini 2.5 Flash 41 for both Smartbook and the reference summarization baseline.”
• mmBERT: A Modern Multilingual Encoder with Annealed Language Learning (2025, unknown): “MM BERT improves over XLM-R across the board, and even beats models like OpenAI’s o3 (OpenAI, 2025) and Google’s Gemini 2.5 Pro (Comanici et al., 2025) on low-resource languages.”