Core Insights - The Shanghai Artificial Intelligence Laboratory has launched the Intern-S1-Pro, a trillion-parameter scientific multimodal model, marking a significant advancement in AI4S from a "tool revolution" to a "revolutionary tool" driving scientific discovery [1][2] - Intern-S1-Pro is the largest scientific multimodal model in the global open-source community, demonstrating superior capabilities in complex scientific reasoning and logic, achieving gold medal-level performance in international assessments [1][3] Model Architecture and Innovations - Intern-S1-Pro utilizes a mixture of experts (MoE) architecture with 512 experts and a total of 1 trillion parameters, activating only 8 experts and 22 billion parameters per call, enhancing both general and scientific capabilities [1][5] - The model incorporates innovations such as Fourier Position Encoding and a restructured temporal encoder, enabling a unified understanding of various scientific signals from micro to macro scales [5][6] - A new routing mechanism has been developed to improve training efficiency and resource allocation, addressing traditional challenges in large-scale model training [6][7] Performance Metrics - Intern-S1-Pro has achieved outstanding results across various benchmarks, including: - 55.5 in comprehensive scientific reasoning (SciReasoner) - 72.8 in material property assessment (Mat Bench) - 93.1 in mathematical reasoning (AIME-2025) [3][4][8] - The model excels in both general and specialized scientific tasks, maintaining a leading position in the open-source model landscape [9][10] Applications and Capabilities - Intern-S1-Pro demonstrates advanced capabilities in understanding complex molecular structures and scientific data, extending its application from micro-level tasks to macro-level analyses [9][10] - The model supports a wide range of scientific disciplines, including chemistry, materials science, biology, earth science, and physics, with over 100 specialized sub-tasks [8][9] Open Source and Community Engagement - The launch of Intern-S1-Pro is part of a broader initiative to create an open and collaborative AGI4S infrastructure, aimed at lowering barriers for global research teams [2][13] - The Shanghai AI Laboratory has established a comprehensive open-source tool ecosystem for model development and application, fostering an active community of developers [12][13]

2月4日，上海人工智能实验室开源基于 "通专融合"技术架构SAGE 打造的 万亿参数 科学多模态大模型 Intern-S1 - Pro ，为AI4S从"工具革命"的 1.0阶段迈向以"革命 的工具"驱动科学发现的2.0时代，提供创新的系统性开源基座。 作为当前全球开源社区中参数规模最大的科学多模态模型，Intern-S1-Pro的核心科 学能力实现了质的跃升，高难度综合学科评测稳居AI4S领域国际领先水平，复杂数理 逻辑推理能力达奥赛金牌水平，面向真实科研流程的智能体能力位居开源模型第一梯 队。 此次发布的Intern-S1-Pro是通过SAGE实现 "可深度专业化通用模型" 的关键实 践。该模型基于混合专家架构（MoE），共拥有512个专家，总参数达1T，每次调用 仅激活8个专家、22B参数。其通用能力和科学能力协同演进，并在底层架构实现了 两大核心突破：在SAGE的基础模型层，通过引入 傅里叶位置编码 ① 并 重构时序编 码器 ，赋予模型统一理解从微观生命信号到宏观宇宙波动的"物理直觉"；通过高效 路由机制 ，系统攻克了训练万亿参数MoE模型在稳定性与算力效率上的瓶颈，为超大 规模模型的训练提供了关键的工 ...