The Gemini 1.5 family of models represents a significant advancement in multimodal understanding, capable of processing up to 10 million tokens of context, including long documents, hours of video and audio. The family includes two models: Gemini 1.5 Pro, which outperforms previous versions on most capabilities and benchmarks, and Gemini 1.5 Flash, a more efficient variant with minimal quality loss. These models achieve near-perfect recall on long-context retrieval tasks across modalities, improve state-of-the-art performance in long-document QA, long-video QA, and long-context ASR, and match or surpass Gemini 1.0 Ultra in various benchmarks. Gemini 1.5 models demonstrate exceptional long-context capabilities, with near-perfect retrieval up to 10 million tokens, surpassing existing models like Claude 3.0 and GPT-4 Turbo. Real-world applications show significant time savings across 10 job categories, and the models can learn to translate English to Kalamang, a language with fewer than 200 speakers, at a level comparable to a human learner. Gemini 1.5 Pro and Flash show improved performance across reasoning, coding, vision, and video benchmarks, with Gemini 1.5 Pro outperforming previous models in core capabilities like math, science, and multilinguality. The models also excel in multimodal reasoning, code generation, and function calling. Gemini 1.5 Pro achieves state-of-the-art results on several multimodal benchmarks, including AI2D, MathVista, and EgoSchema. The models are designed for high efficiency and low latency, with Gemini 1.5 Flash being the fastest for all languages tested. Training infrastructure includes Google's TPUv4 accelerators and a diverse multimodal dataset. Evaluation results show significant improvements in long-context understanding, with Gemini 1.5 Pro achieving near-perfect recall up to 10 million tokens across text, video, and audio. The models also demonstrate strong performance in realistic tasks like in-context learning, translation, and speech recognition, with Gemini 1.5 Pro achieving 100% recall in translating from English to Kalamang. These results highlight the models' ability to handle complex, long-context tasks and their potential in supporting endangered languages and facilitating cross-linguistic communication.The Gemini 1.5 family of models represents a significant advancement in multimodal understanding, capable of processing up to 10 million tokens of context, including long documents, hours of video and audio. The family includes two models: Gemini 1.5 Pro, which outperforms previous versions on most capabilities and benchmarks, and Gemini 1.5 Flash, a more efficient variant with minimal quality loss. These models achieve near-perfect recall on long-context retrieval tasks across modalities, improve state-of-the-art performance in long-document QA, long-video QA, and long-context ASR, and match or surpass Gemini 1.0 Ultra in various benchmarks. Gemini 1.5 models demonstrate exceptional long-context capabilities, with near-perfect retrieval up to 10 million tokens, surpassing existing models like Claude 3.0 and GPT-4 Turbo. Real-world applications show significant time savings across 10 job categories, and the models can learn to translate English to Kalamang, a language with fewer than 200 speakers, at a level comparable to a human learner. Gemini 1.5 Pro and Flash show improved performance across reasoning, coding, vision, and video benchmarks, with Gemini 1.5 Pro outperforming previous models in core capabilities like math, science, and multilinguality. The models also excel in multimodal reasoning, code generation, and function calling. Gemini 1.5 Pro achieves state-of-the-art results on several multimodal benchmarks, including AI2D, MathVista, and EgoSchema. The models are designed for high efficiency and low latency, with Gemini 1.5 Flash being the fastest for all languages tested. Training infrastructure includes Google's TPUv4 accelerators and a diverse multimodal dataset. Evaluation results show significant improvements in long-context understanding, with Gemini 1.5 Pro achieving near-perfect recall up to 10 million tokens across text, video, and audio. The models also demonstrate strong performance in realistic tasks like in-context learning, translation, and speech recognition, with Gemini 1.5 Pro achieving 100% recall in translating from English to Kalamang. These results highlight the models' ability to handle complex, long-context tasks and their potential in supporting endangered languages and facilitating cross-linguistic communication.