Intern-S2-Preview - Shanghai AI Lab's Open-Source Scientific Multimodal Large Model

What is Intern-S2-Preview

Intern-S2-Preview is the preview version of Shanghai Artificial Intelligence Laboratory's new-generation open-source Intern scientific multimodal large model, achieving scientific capabilities comparable to trillion-parameter models with a 35B parameter scale. Driven by "general-purpose and specialized integration" full-chain training and reinforcement learning, it is the first open-source general-purpose large model to achieve material crystal structure generation, scoring 57.26 on the MolecularIQ benchmark, with a crystal structure generation pass rate exceeding 40%, significantly surpassing mainstream closed-source models. It also supports complex scientific reasoning, biological multi-omics understanding, and agent task execution. Based on the Ascend Atlas 900 A3 supernode, it achieves algorithm-system-computing power collaborative optimization, providing efficient, low-barrier AI infrastructure for scientific research innovation.

Intern-S2-Preview's Main Features

• Scientific multimodal understanding: Supports multi-omics sequence analysis, biomedical microscopy image question answering, biomolecular instruction understanding, molecular structure reasoning, scientific multimodal tasks, remote sensing perception and reasoning, and comprehensive scientific reasoning.
• Material crystal structure generation: First in an open-source general-purpose large model to achieve material crystal structure generation, introducing a real-number prediction module, supporting high-precision coordinate regression and molecular structure spatial modeling.
• Mathematics and complex reasoning: Covers tasks such as International Mathematical Olympiad problem solving, multimodal mathematical reasoning, and high school mathematics competitions, supporting chain-of-thought folding and highly compressed efficient reasoning.
• Long-range text and multimodal reasoning: Supports long-range pure text and multimodal reasoning, with the ability to process ultra-long scientific literature and complex cross-modal information.
• Scientific chart and data question answering: Provides professional question answering and analysis for scientific charts and data visualization content.
• Scientific code generation: Performs excellently in scientific programming and algorithm solving scenarios, efficiently supporting scientific computing, algorithm development, and research script writing.
• General-purpose agent task execution: Possesses complex task planning and autonomous execution capabilities for scientific agent interaction, scientific competition and research agent task execution, OpenClaw coding agents, and software engineering agents.
• Instruction following and code production: Supports high-precision instruction following and general-purpose code production tasks.

Intern-S2-Preview's Technical Principles

• General-purpose and specialized integration full-chain training: Expands hundreds of specialized scientific tasks from pre-training to the reinforcement learning stage, equipping each task with high-quality data and training strategies from pre-training to post-training, achieving multi-task integrated training and enabling synergistic effects among different tasks.
• Task Scaling mechanism: Achieves capability Scaling by increasing task difficulty and enriching task diversity. When a large number of high-difficulty scientific tasks are unified and integrated for training, a 35B parameter model can reach the performance level of trillion-parameter models on multiple scientific tasks.
• Reinforcement learning-driven scientific reasoning: Extends the RL training step length and introduces graduate-level disciplinary reasoning problems; guides the model to use chain-of-thought to complete professional tasks such as biological multi-omics understanding, relying on the generalization advantages of chain-of-thought to achieve performance comparable to large models with small parameters.
• Data thought density optimization (IQPT): Guided by the Intelligence Quality per token concept, it explores innovative algorithms such as chain-of-thought folding, and by building a data thinking density lever, it matches 8x-parameter models with the same unit intelligence in mathematical reasoning.
• Molecular Structure Spatial Modeling Technology: Introduces Fourier Position Encoding (FoPE), reconstructs the temporal encoder, and adds a real-number prediction module to achieve high-precision coordinate regression, completing material crystal structure generation in an open-source general large model for the first time.
• Integrated Training and Inference Architecture: Based on the XTuner training framework and the LMDeploy inference engine, it introduces a shared MTP weight computation method, reduces inconsistency between training and inference stages, and improves draft token acceptance rate and generation effectiveness.
• Ascend Compute Collaborative Optimization: Based on the Ascend Atlas 900 A3 super node, it introduces memory optimization techniques such as SP, chunk loss, and activation offload into the training framework; optimizes data chunking and host-device interaction for variable-length inputs; and by offline simulating the compute proportion of vision and language modules, achieves balanced resource allocation for multimodal long-sequence training.

How to Use Intern-S2-Preview

• Online Chat Experience: Visit the official Shusheng Large Model experience platform to directly conduct multi-turn conversations and test capabilities such as scientific Q&A, molecular structure understanding, and code generation.
• API Service Integration: Call the model via ChatAPI, in the model field fill in intern-s2-preview, with deep thinking mode enabled by default (thinking_mode), suitable for scenarios that need to be integrated into scientific research toolchains or automated workflows.
• Agent Task Execution: If you need to use agent capabilities (such as connecting to harnesses like OpenClaw to execute complex scientific research tasks), it is strongly recommended to keep thinking_mode enabled to ensure the stability of task decomposition, tool invocation, and multi-step decision-making.
• Local Deployment of Open-Source Models: Download model weights from HuggingFace or ModelScope, and perform local deployment with the officially recommended LMDeploy inference engine, supporting efficient inference and multimodal long-sequence processing.
• Scientific Research Fine-Tuning and Training: Based on the XTuner training framework, use open-source weights for domain fine-tuning; the training stage supports multi-token prediction reinforcement learning and shared MTP weight computation, facilitating integrated training and inference iteration.
• Crystal Structure Generation Task: For materials science scenarios, directly call the model for crystal structure generation and molecular coordinate regression, obtaining high-precision spatial coordinate prediction results without relying on diffusion models.

Intern-S2-Preview Project Address

• Shusheng Large Model：https://chat.intern-ai.org.cn/
• HuggingFace：https://huggingface.co/internlm/Intern-S2-Preview
• ModelScope：https://modelscope.cn/models/Shanghai_AI_Laboratory/Intern-S2-Preview

Core Advantages of Intern-S2-Preview

• 35B Parameters Comparable to Trillion-Level Models: With an extremely small parameter scale, it achieves capabilities comparable to trillion-parameter models in multiple core scientific fields, greatly lowering the threshold for scientific research use and deployment costs.
• First Open-Source Crystal Structure Generation: Introduces a real-number prediction module, achieving material crystal structure generation for the first time in an open-source general large model, with a MolecularIQ score of 57.26 and a pass rate exceeding 40%, significantly surpassing closed-source models such as GPT-5.5.
• Leading Scientific Agent Capabilities: Surpasses mainstream closed-source models such as Claude-Haiku-4.5 and GPT5.4-Nano in comprehensive scientific scenario programming and scientific discovery tasks, ranking among the top of the same scale in SciCode and PinchBench evaluations.
• Full-Chain Training Integrating General and Specialized Capabilities: Extend professional scientific tasks from pretraining to the reinforcement learning stage, where multi-task integration creates synergistic effects and avoids the trade-offs in capabilities caused by optimizing a single stage.
• Reinforcement Learning Drives Efficient Reasoning: Relying on chain-of-thought guidance and chain-of-thought folding algorithms, it matches models with 8 times the parameter count at the same unit intelligence in mathematical reasoning, achieving a dual breakthrough in performance and efficiency.
• Ascend Software-Hardware Collaborative Optimization: Based on the Ascend Atlas 900 A3 supernode, it achieves collaborative evolution of algorithms, systems, and computing power, greatly improving training stability and inference efficiency, and validating the value of the domestic computing power ecosystem.
• Integrated Training and Inference Architecture: The XTuner training framework and LMDeploy inference engine are deeply coordinated, sharing MTP weights to reduce training-inference inconsistency, and supporting efficient training and deployment of multimodal long sequences.

Comparison of Intern-S2-Preview with Similar Competitors

Application Scenarios of Intern-S2-Preview

• Molecular Biology and Drug Development: Used for multi-omics sequence analysis, biomolecular instruction understanding, and molecular structure reasoning, assisting in target discovery, compound screening, and research on drug mechanisms of action.
• Innovative Discovery in Materials Science: Directly generates material crystal structures and outputs high-precision spatial coordinates, accelerating the R&D iteration of new semiconductors, catalysts, battery materials, etc.
• Scientific Computing and Algorithm Development: Supports scientific programming, algorithm solving, and research script writing, providing automated code generation for physics simulation, chemistry computation, and bioinformatics processing.
• Biological microscopy images and remote sensing analysis: Provides professional question answering and feature recognition for biological microscopy images, while also supporting remote sensing perception and reasoning, serving the fields of medical imaging and Earth science.
• Mathematics and complex science reasoning: Covers the International Mathematical Olympiad, high school mathematics competitions, and graduate-level disciplinary reasoning, assisting with mathematical proofs, formula derivation, and logic verification.
• Automated execution by scientific research agents: As scientific competition agents, research agents, or OpenClaw coding agents, autonomously completes multi-step research tasks such as literature review, experimental design, and data analysis.