Intern-S2-Preview - Open-source scientific multimodal large model from Shanghai AI Lab

What is Intern-S2-Preview?

Intern-S2-Preview is an open-source preview of the next-generation Shusheng Scientific Multimodal Large Model from the Shanghai Artificial Intelligence Laboratory. With only 35 billion parameters, it achieves scientific capabilities comparable to models with trillions of parameters. Driven by a comprehensive training and reinforcement learning approach, the model is the first open-source general-purpose large model to generate material crystal structures, achieving a MolecularIQ score of 57.26 and a crystal structure generation success rate exceeding 40%, significantly surpassing mainstream closed-source models. It also supports complex scientific reasoning, multi-omics understanding of biology, and intelligent agent task execution. Based on the Ascend Atlas 900 A3 supernode, it achieves algorithm-system-computing power collaborative optimization, providing efficient and low-barrier AI infrastructure for scientific research and innovation.

Main functions of Intern-S2-Preview

• Scientific Multimodal UnderstandingIt supports multi-omics sequence analysis, biological microscopic image question answering, biomolecular instruction understanding, molecular structure reasoning, scientific multimodal tasks, remote sensing perception and reasoning, and comprehensive scientific reasoning.
• Material crystal structure generationThis is the first time that material crystal structure generation has been achieved in an open-source general-purpose large model. A real number prediction module has been introduced to support high-precision coordinate regression and molecular structure spatial modeling.
• Mathematics and Complex ReasoningIt covers tasks such as solving problems in the International Mathematical Olympiad, multimodal mathematical reasoning, and high school mathematics competitions, and supports efficient reasoning through the folding and extreme compression of thought chains.
• Long-range text and multimodal reasoningIt supports long-range plain text and multimodal reasoning, and has the ability to process ultra-long scientific documents and complex cross-modal information.
• Scientific charts and data Q&AWe provide professional Q&A and analysis for scientific charts and data visualizations.
• Scientific code generationIt performs excellently in scientific programming and algorithm solving scenarios, and can efficiently support scientific computing, algorithm development and scientific script writing.
• General intelligent agent task executionIt possesses the ability to plan and autonomously execute complex tasks such as scientific intelligent agent interaction, scientific competition and research intelligent agent task execution, OpenClaw coded intelligent agents, and software engineering intelligent agents.
• Instruction following and code productionSupports high-precision instruction following and general code production tasks.

Technical Principles of Intern-S2-Preview

• Integrated training across multiple disciplinesIt extends hundreds of professional scientific tasks from pre-training to reinforcement learning, equipping each task with high-quality data and training strategies from pre-training to post-training, enabling multi-task fusion training and creating a synergistic effect that promotes mutual development among different tasks.
• Task Scaling MechanismBy increasing task difficulty and enriching task diversity, scaling is achieved. When a large number of highly challenging scientific tasks are integrated and trained in a unified manner, the 35B parameter model can achieve the performance level of a trillion parameter model on multiple scientific tasks.
• Reinforcement learning drives scientific reasoningExtend the training step size of RL and introduce graduate-level subject reasoning problems; guide the model to complete professional tasks such as understanding biological multi-omics by using the thinking chain, and achieve performance comparable to large models with small parameters by relying on the generalization advantage of the thinking chain.
• Data-driven thinking density optimization (IQPT)Guided by the concept of Intelligence Quality per token, we explore innovative algorithms such as mind chain folding. By building a data thinking density lever, we can achieve the same level of intelligence in mathematical reasoning as a model with 8 times the number of parameters.
• Molecular structure spatial modeling technologyThe invention introduces Fourier position coding (FoPE), reconstructs the temporal encoder, and adds a real number prediction module to achieve high-precision coordinate regression. For the first time, it completes the generation of material crystal structure in an open-source general-purpose large model.
• Integrated training and push architectureBased on the XTuner training framework and the LMDeploy inference engine, a shared MTP weight calculation method is introduced to reduce inconsistencies between the training and inference stages and improve the draft token acceptance rate and generation effectiveness.
• Ascend computing power collaborative optimizationBased on the Ascend Atlas 900 A3 supernode, memory optimization techniques such as SP, chunk loss, and activation offload are introduced into the training framework; data block and host device interaction are optimized for variable-length input; and balanced resource allocation for multimodal long sequence training is achieved by offline simulation of the computing power ratio of vision and language modules.

How to use Intern-S2-Preview

• Online conversation experienceVisit the official experience platform of Shusheng Big Model to directly engage in multi-round dialogues and test capabilities such as scientific question answering, molecular structure understanding, and code generation.
• API service access: Call the model via ChatAPI, in model Field filling intern-s2-previewDeep thinking mode is enabled by default.thinking_modeIt is suitable for scenarios that need to be integrated into scientific research toolchains or automated workflows.
• Intelligent agent task executionIf you need to use agent capabilities (such as connecting to harnesses like OpenClaw to perform complex research tasks), it is strongly recommended to maintain... thinking_mode Enabled to ensure the stability of task breakdown, tool calls, and multi-step decision-making.
• Local deployment of open source modelsDownload model weights from HuggingFace or ModelScope and deploy them locally using the officially recommended LMDeploy inference engine, supporting efficient inference and multimodal long sequence processing.
• Research fine-tuning and trainingBased on the XTuner training framework, it utilizes open-source weights for domain fine-tuning; during the training phase, it supports multi-token prediction reinforcement learning and shared MTP weight calculation, facilitating integrated training and induction iteration.
• Crystal structure generation taskFor materials science scenarios, the model can be directly called to generate crystal structures and regress molecular coordinates, and high-precision spatial coordinate prediction results can be obtained without relying on diffusion models.

Intern-S2-Preview project address

• Scholar Modelhttps://chat.intern-ai.org.cn/
• HuggingFacehttps://huggingface.co/internlm/Intern-S2-Preview
• ModelScope: https://modelscope.cn/models/Shanghai_AI_Laboratory/Intern-S2-Preview

The core advantages of Intern-S2-Preview

• 35B parameters rival trillion-level modelsIt enables the use of trillion-parameter models in multiple core scientific fields with extremely small parameter scale, significantly reducing the threshold for scientific research and deployment costs.
• First open source crystal structure generationThe introduction of a real number prediction module enables the generation of material crystal structures for the first time in an open-source general-purpose large model. The MolecularIQ score is 57.26, with a pass rate of over 40%, significantly surpassing closed-source models such as GPT-5.5.
• Leading scientific intelligent agent capabilitiesIt surpasses mainstream closed-source models such as Claude-Haiku-4.5 and GPT5.4-Nano in comprehensive scientific programming and scientific discovery tasks, and ranks among the top in its class in SciCode and PinchBench evaluations.
• Integrated training across multiple disciplinesExtending professional scientific tasks from pre-training to reinforcement learning stage, integrating multiple tasks to form a synergistic effect, and avoiding the ebb and flow of optimization capabilities in a single stage.
• Reinforcement learning drives efficient reasoningBased on the thinking chain guidance and thinking chain folding algorithm, it achieves a breakthrough in both performance and efficiency in mathematical reasoning by using the same unit of intelligence to match a model with 8 times the number of parameters.
• Ascend hardware and software co-optimizationBased on the Ascend Atlas 900 A3 supernode, the algorithm, system, and computing power are coordinated and evolved, which greatly improves training stability and inference efficiency, and verifies the value of the domestic computing power ecosystem.
• Integrated training and push architectureThe XTuner training framework and the LMDeploy inference engine work together in a deep collaboration, sharing MTP weights to reduce training-inference inconsistencies and supporting efficient training and deployment of multimodal long sequences.

Comparison of Intern-S2-Preview with similar products

Application scenarios of Intern-S2-Preview

• Molecular biology and drug developmentIt is used for multi-omics sequence analysis, understanding of biomolecular instructions and reasoning about molecular structures, and to assist in target discovery, compound screening and research on drug action mechanisms.
• Innovative discoveries in materials scienceIt can directly generate the crystal structure of materials and output high-precision spatial coordinates, accelerating the research and development of new semiconductors, catalysts, battery materials, etc.
• Scientific computing and algorithm developmentIt supports scientific programming, algorithm solving, and scientific script writing, and provides automated code generation for physical simulation, chemical calculation, and bioinformatics processing.
• Biological Microscopic Images and Remote Sensing AnalysisIt provides professional question answering and feature recognition for biological microscopic images, while also supporting remote sensing perception and reasoning, serving the fields of medical imaging and earth science.
• Mathematics and Complex Scientific ReasoningIt covers subject-specific reasoning for international mathematics Olympiads, high school mathematics competitions, and postgraduate level, assisting in mathematical proofs, formula derivations, and logical verifications.
• Automated execution of scientific research intelligent agentsAs a science competition agent, research agent, or OpenClaw coding agent, it can autonomously complete multi-step scientific research tasks such as literature review, experimental design, and data analysis.