Source-linked AI summary
Intern-S1-Pro: Scientific Multimodal Foundation Model at Trillion Scale
Yicheng Zou, Dongsheng Zhu, Lin Zhu, Tong Zhu, Yunhua Zhou, Peiheng Zhou, Xinyu Zhou, Dongzhan Zhou, Zhiwang Zhou, Yuhao Zhou, Bowen Zhou, Zhanping Zhong, Zhijie Zhong, Haiteng Zhao, Penghao Zhao, Xiaomeng Zhao, Zhiyuan Zhao, Yechen Zhang, Jin Zhang, Wenwei Zhang, Hongjie Zhang, Zhuo Zhang, Wenlong Zhang, Bo Zhang, Chao Zhang, Chen Zhang, Yuhang Zang, Fei Yuan, Jiakang Yuan, Jiashuo Yu, Jinhui Yin, Haochen Ye, Qian Yao, Bowen Yang, Danni Yang, Kaichen Yang, Ziang Yan, Jun Xu, Yicheng Xu, Wanghan Xu, Xuenan Xu, Chao Xu, Ruiliang Xu, Shuhao Xing, Long Xing, Xinchen Xie, Ling-I Wu, Zijian Wu, Zhenyu Wu, Lijun Wu, Yue Wu, Jianyu Wu, Wen Wu, Fan Wu, Xilin Wei, Qi Wei, Bingli Wang, Rui Wang, Ziyi Wang, Zun Wang, Yi Wang, Haomin Wang, Yizhou Wang, Lintao Wang, Yiheng Wang, Longjiang Wang, Bin Wang, Jian Tong, Zhongbo Tian, Huanze Tang, Chen Tang, Shixiang Tang, Yu Sun, Qiushi Sun, Xuerui Su, Qisheng Su, Chenlin Su, Demin Song, Jin Shi, Fukai Shang, Yuchen Ren, Pengli Ren, Xiaoye Qu, Yuan Qu, Jiantao Qiu, Yu Qiao, Biqing Qi, Runyu Peng, Tianshuo Peng, Jiahui Peng, Qizhi Pei, Zhuoshi Pan, Linke Ouyang, Wenchang Ning, Yichuan Ma, Zerun Ma, Ningsheng Ma, Runyuan Ma, Chengqi Lyu, Haijun Lv, Han Lv, Lindong Lu, Kuikun Liu, Jiangning Liu, Yuhong Liu, Kai Liu, Hongwei Liu, Zhoumianze Liu, Mengjie Liu, Ziyu Liu, Wenran Liu, Yang Liu, Liwei Liu, Kaiwen Liu, Junyao Lin, Junming Lin, Tianyang Lin, Dahua Lin, Jianze Liang, Linyang Li, Peiji Li, Zonglin Li, Zehao Li, Pengze Li, Guoyan Li, Lingkai Kong, Linglin Jing, Zhenjiang Jin, Feifei Jiang, Qian Jiang, Junhao Huang, Zixian Huang, Haian Huang, Zhouqi Hua, Ermo Hua, Han Hu, Linfeng Hou, Yinan He, Conghui He, Tianyao He, Xu Guo, Qipeng Guo, Aijia Guo, Yuzhe Gu, Lixin Gu, Jingyang Gong, Qiming Ge, Jiaye Ge, Songyang Gao, Jianfei Gao, Xinyu Fang, Caihua fan, Yue Fan, Yanhui Duan, Zichen Ding, Shengyuan Ding, Ning Ding, Xuanlang Dai, Erfei Cui, Ganqu Cui, Pei Chu, Tao Chu, Guangran Cheng, Yu Cheng, Kai Chen, Yongkang Chen, Chiyu Chen, Guanzhou Chen, Qiaosheng Chen, Sitao Chen, Xin Chen, Haojiong Chen, Yicheng Chen, Weihan Cao, Yuhang Cao, Qinglong Cao, Lei Bai
TL;DR
Scientific foundation models must cover diverse, specialized domains while preserving multimodal and reasoning capabilities. Intern-S1-Pro addresses this gap with trillion-parameter scaling, joint general-specialized training, expert and data strategies, and infrastructure optimized for large-scale RL. It reports state-of-the-art scientific benchmark performance and stronger scientific reasoning than leading closed-source models, while extending coverage across scientific modalities and tasks.
Problem
Scientific domains contain diverse specialized languages, long-tailed knowledge, and complex multimodal data, while high-quality scientific image-text pairs remain difficult to obtain.
Method
Intern-S1-Pro combines trillion-parameter expert expansion with Grouped Routing, continued pre-training on 6T multimodal tokens, and a specialized scientific image-captioning pipeline.
Results
Intern-S1-Pro achieves state-of-the-art performance across a wide range of scientific benchmarks and outperforms proprietary models on multiple scientific evaluations.
Takeaways & Limitations
The model functions as a broad scientific generalist with expanded specialized coverage across chemistry, materials, life sciences, earth sciences, and additional time-series domains.
Takeaways & Limitations
AgentCompass is an agent evaluation framework developed by the authors that will be released in the future.
Abstract
from arXiv · showhide
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.
1. Introduction
Intern-S1-Pro scales scientific multimodal modeling to one trillion parameters to combine stronger general intelligence with broader scientific expertise. Its joint training and infrastructure support target stable, efficient scaling across general and specialized capabilities.
- Scientific diversity and long-tailed frontier knowledge motivate scaling model size beyond the requirements of narrower language settings.
- Intern-S1-Pro is the first one-trillion-parameter scientific multimodal foundation model, enhancing general and scientific capabilities.
- Its scientific expertise covers over 100 specialized tasks spanning chemistry, materials, life sciences, and earth sciences.
- Joint training on general and specific tasks enables Intern-S1-Pro to outperform specialized models in several scientific tasks.
- XTuner and LMDeploy co-design scales the model to 4× Intern-S1 with only a ∼20% reduction in training efficiency while supporting trillion-parameter RL training.
2. Architecture
The architecture combines expert expansion and routing mechanisms for stable, balanced trillion-scale MoE training with multimodal and scientific encoders. It also adapts representation and temporal processing to preserve information across visual, spectral, and time-series inputs.
- 2.1. Group Routing: Grouped Routing distributes expanded experts into groups aligned with pretrained Top-1 or Top-2 experts, improving training stability after expansion.
- 2.1. Group Routing: Traditional Top-k routing can create cross-device load imbalance, reducing efficiency and increasing out-of-memory risk during expert-parallel training.
- 2.1. Group Routing: Grouped Router replaces traditional Top-K routing to achieve absolute load balancing across devices under 8-way expert parallelism.
- 2.2. Straight-Through Estimator for Sparse Expert Routing: The Straight-Through Estimator preserves sparse Top-K selection during the forward pass while allowing gradients to flow through softmax routing during backpropagation.
- 2.4. FoPE: FoPE represents each dimension with Fourier frequency components to separate information, mitigate spectral damage, and clip inadequately trained components.
- 2.5. Time-series Encoder: Intern-S1-Pro expands time-series coverage from astronomy, geoscience, and neuroscience to physiological signal analysis and bioacoustics.
- 2.5. Time-series Encoder: The time-series module uses adaptive subsampling and an encoder to capture local patch dynamics and long-range dependencies across segments.
3. Pre-training
Intern-S1-Pro improves scientific multimodal pre-training through dense, aligned scientific image-text data and strategies that reduce conflicts between scientific and general data. Its pipeline transforms PDF figures into high-quality captions while data and prompt upgrades support broader, more stable multimodal performance.
- Scientific multimodal data: 6T tokens of image-text and text data support Intern-S1-Pro’s continued pre-training, including captions tailored for scientific images.Scientific images require greater content accuracy and attention to detail than natural images.
- Scientific multimodal data: Scientific image captions from literature are often brief and weakly aligned, motivating a dedicated pipeline for high-quality image-text pairs.The pipeline targets stronger understanding of scientific visual content.
- Caption pipeline: PDF corpora provide dense scientific visual content, and the pipeline extracts sub-figures to construct strongly aligned image-text training data.It uses figures, formulas, tables, and other high-information-density scientific visuals as source material.
- Caption pipeline: MinerU2.5 detects and crops scientific visual elements, while topic-based model routing generates domain-specific or general-purpose dense captions.InternVL3.5-241B handles scientific sub-images, whereas CapRL-32B processes non-scientific sub-images.
- Caption pipeline: Multi-template prompting and a 0.5B-parameter text quality discriminator increase linguistic diversity while filtering garbled, repetitive, and low-information content.The pipeline was deployed across life sciences, chemistry, earth sciences, and materials science PDF corpora.
- Data integration: Template construction, task-form transformation, prompt diversification, and system prompt isolation mitigate scientific-general data conflicts and improve multimodal stability.These strategies convert structured scientific inputs, diversify instructions, and create separate contextual environments for scientific and general data.
4. Post-Training
Post-training addresses the memory and stability challenges of trillion-parameter sparse-MoE reinforcement learning. The approach combines precision-sensitive engine alignment, FP8-aware optimization, and discrepancy handling so mixed-precision RL matches BF16 behavior in practice.
- Motivation: Trillion-parameter sparse-MoE reinforcement learning creates substantial memory pressure because Intern-S1-Pro has 4× Intern-S1’s expert count.Comparable activated parameters do not remove the memory burden from expert-layer parameters and activations.
- Engine consistency: Operator-level comparison between LMDeploy and XTuner identifies RMSNorm, router softmax, and positional embeddings as numerically sensitive components.Reducing precision gaps in these kernels makes the rollout distribution more faithfully reflected during training.
- Mixed-precision RL: FP8 mixed-precision RL preserves most memory and throughput benefits while avoiding unnecessary degradation in sensitive computation-graph components.The method modifies the REINFORCE objective with dual importance sampling ratios.
- Discrepancy control: The masking function suppresses tokens with excessively large training–rollout discrepancies during optimization.This addresses distribution mismatch between training and rollout behavior.
- Advantage estimation: A leave-one-out baseline estimates advantages across sampled responses, applying each response’s advantage consistently to all its tokens.The sequence-level reward is denoted by R_i, and the same estimated advantage is used across tokens t within response y_i.
- Results: FP8 mixed-precision RL matches BF16 training behavior in practice, as shown by average accuracy curves across multiple validation benchmarks.The comparison uses a 30B MoE model during RL optimization.
5. Evaluation
Intern-S1-Pro is evaluated across scientific and general-purpose benchmarks in text-only and multimodal settings, using thinking and non-thinking configurations. It shows strong scientific and general performance, with gains over prior and specialized models and improvements on scientific time-series tasks.
- Evaluation setup: Intern-S1-Pro is evaluated across scientific and general-purpose benchmarks covering text-only and multimodal settings.The evaluation uses OpenCompass, VLMEvalKit, and Agent-Compass, with benchmark-specific thinking or non-thinking configurations.
- Scientific benchmarks: 55.5 on SciReasoner exceeds Gemini-3-Pro’s 14.7 and GPT-5.2’s 13.6, while Intern-S1-Pro leads several other scientific benchmarks.Reported leading scores include 74.8 on SmolInstruct, 72.8 on MatBench, 48.8 on Mol-Instructions, 52.5 on Biology-Instruction, and 52.8 on XLRS-Bench.
- General benchmarks: 93.1 on AIME-2025 and 86.6 on MMLU-Pro show strong general-task performance matching or exceeding strong open-source models.The comparison specifically names Qwen3-VL-235B-Thinking as a strong open-source reference.
- Improvements over Intern-S1: AIME-2025 improves from 86.0 to 93.1 and MMLU-Pro from 83.5 to 86.6 relative to Intern-S1, alongside expanded scientific coverage and enhanced agent results.The reported agent scores are 77.4 on GAIA (Text-Only), 80.9 on τ2-Bench, and 93.6 on ScreenSpot V2.
- Time-series results: 99.5 F1 on SciTS task EAU01 and improvements on BIU03 and PHU04 support the dedicated time-series module’s effectiveness.The passage attributes these gains to dynamic subsampling and a dedicated time-series encoder for capturing complex temporal dynamics.
- Biology case study: Intern-S1-Pro achieves better comprehensive performance than the specialized Biology-Instruction model, including 78.14 versus 2.57 on Protein-Fluorescence.The models used the same underlying dataset, with only more fluent text expression added for Intern-S1-Pro; Protein-FunctionEC scores were 72.70 versus 19.79.
6. Conclusion
Intern-S1-Pro combines trillion-parameter expert expansion, Grouped Routing, large-scale multimodal pre-training, and scientific captioning to advance scientific multimodal modeling. Evaluations report state-of-the-art performance across broad scientific benchmarks, with robust reasoning and deep domain knowledge.
- Intern-S1-Pro scales to a trillion parameters through expert expansion and Grouped Routing, improving load balancing and training stability.The design mitigates expert homogenization and training instability in large-scale MoE models.
- 6T tokens of high-quality multimodal data strengthen the model’s scientific understanding through continued pre-training.
- Alignment-focused captions for scientific figures improve interpretation of complex scientific visual content.
- Intern-S1-Pro achieves state-of-the-art performance across a wide range of scientific benchmarks, demonstrating robust reasoning and deep domain knowledge.