We introduce Intern-S1-Pro, the first one-trillion-parameter scientific multimodal foundation model. Scaling to this unprecedented size, the model delivers a comprehensive enhancement across both general and scientific domains. Beyond stronger reasoning and image-text understanding capabilities, its intelligence is augmented with advanced agent capabilities. Simultaneously, its scientific expertise has been vastly expanded to master over 100 specialized tasks across critical science fields, including chemistry, materials, life sciences, and earth sciences. Achieving this massive scale is made possible by the robust infrastructure support of XTuner and LMDeploy, which facilitates highly efficient Reinforcement Learning (RL) training at the 1-trillion parameter level while ensuring strict precision consistency between training and inference. By seamlessly integrating these advancements, Intern-S1-Pro further fortifies the fusion of general and specialized intelligence, working as a Specializable Generalist, demonstrating its position in the top tier of open-source models for general capabilities, while outperforming proprietary models in the depth of specialized scientific tasks.
核心贡献 · Key contributions
首个万亿参数科学多模态基础模型 Intern-S1-Pro，在 100 多项专业科学任务上达到最先进水平。 First one-trillion-parameter scientific multimodal foundation model, Intern-S1-Pro, achieving state-of-the-art performance across 100+ specialized scientific tasks.
分组路由机制实现设备间绝对负载均衡，消除显存溢出风险并提升大规模 MoE 训练稳定性。 Grouped Routing mechanism ensures absolute load balancing across devices, eliminating OOM risks and improving training stability in large-scale MoE.
直通估计器用于路由器梯度估计，加速扩展 MoE 架构中专家嵌入的学习。 Straight-Through Estimator (STE) for router gradient estimation accelerates expert embedding learning in expanded MoE architectures.
傅里叶位置编码同时捕获离散词元和连续波特征，适用于多模态物理信号。 Fourier Position Encoding (FoPE) captures both discrete token and continuous wave characteristics for multimodal physical signals.
面向科学图像的专用描述流水线从 PDF 语料生成 2700 亿词元的高质量图文对。 Dedicated caption pipeline for scientific images generates 270B tokens of high-quality image-text pairs from PDF corpora.
FP8 混合精度强化学习训练结合回放路由器重放和重要性采样，在万亿参数规模下匹配 BF16 性能。 FP8 mixed-precision RL training with rollout router replay and importance sampling matches BF16 performance at trillion-parameter scale.
局限 · Limitations
评估限于 100 多项科学任务，更广泛的科学领域可能未被覆盖。 Evaluation limited to 100+ scientific tasks; broader scientific domains may not be covered.
训练数据主要来自英文和中文来源，未探索多语言科学数据。 Training data primarily from English and Chinese sources; multilingual scientific data not explored.
分组路由可能导致初始专家同质化，需要若干训练步数才能分化。 Grouped Routing may cause initial expert homogenization, requiring several training steps for differentiation.
FP8 量化仅应用于专家线性层，其他组件保持更高精度。 FP8 quantization applied only to expert linear layers; other components remain in higher precision.
智能体能力在有限基准（GAIA、Tau2-Bench、ScreenSpot V2）上评估，未测试实际部署。 Agent capabilities evaluated on limited benchmarks (GAIA, Tau2-Bench, ScreenSpot V2); real-world deployment not tested.
论文章节 · Sections（共 18）
摘要Abstract
1 引言1 Introduction
2 架构2 Architecture
2.1 分组路由2.1 Group Routing
2.2 稀疏专家路由的直通估计器2.2 Straight-Through Estimator for Sparse Expert Routing
2.3 视觉编码器2.3 Vision Encoder
2.4 FoPE2.4 FoPE
2.5 时间序列编码器2.5 Time-series Encoder
3 预训练3 Pre-training
3.1 标题生成流程3.1 Caption Pipeline
3.2 解决科学与文本数据冲突3.2 Resolving conflicts between the scientific and textual data