InternLM 发布 Intern-S2 系列科学多模态模型

InternLM; Intern-S1

Repository navigation

NotificationsYou must be signed in to change notification settings; Fork54 (54); Star866 (866)You must be signed in to star a repository

A Scientific Multimodal Foundation Model

Repository files navigation

👋 join us on Discord and WeChat

Introduction

We introduce Intern-S2-397B, our most capable multimodal foundation model for scientific intelligence and long-horizon agents. Intern-S2-397B scales along three critical dimensions: pre-training, reinforcement-learning task coverage, and interactive agent environments. By combining a new vision-language pre-training paradigm with large-scale multi-task reinforcement learning and long-horizon agent reinforcement learning, Intern-S2-397B delivers a step change in general reasoning, scientific problem solving, and agentic capabilities.

Features

New pre-training paradigm. Via visual pre-training, Intern-S2-397B learns directly from raw pages of scientific literature, jointly modeling symbolic semantics and visual relationships in a shared representation space without intermediate parsing. This preserves text-visual correspondence, strengthens spatial and visual reasoning, and improves data efficiency; Scientific modality reasoning and generation. By scaling diverse scientific reinforcement-learning tasks across more than 20 domains and training them jointly, Intern-S2-397B achieves leading general-reasoning performance among open-source models and strong results in specialized scientific tasks such as biomolecular interaction design and material structure generation; General and scientific long-horizon agents. By connecting multiple agent frameworks to large-scale sandboxed environments for black-box agentic reinforcement learning, Intern-S2-397B improves generalization and raises the capability ceiling for long-horizon tasks in both general and scientific domains

We introduce Intern-S2-Preview-397B, the preview release of Intern-S2-397B. Intern-S2-Preview-397B scales along three critical dimensions: pre-training, reinforcement-learning task coverage, and interactive agent environments. By combining a new vision-language pre-training paradigm with large-scale multi-task reinforcement learning and long-horizon agent reinforcement learning, Intern-S2-Preview-397B delivers a step change in general reasoning, scientific problem solving, and agentic capabilities.

Features

New pre-training paradigm. Via visual pre-training, Intern-S2-Preview-397B learns directly from raw pages of scientific literature, jointly modeling symbolic semantics and visual relationships in a shared representation space without intermediate parsing. This preserves text-visual correspondence, strengthens spatial and visual reasoning, and improves data efficiency; Scientific modality reasoning and generation. By scaling diverse scientific reinforcement-learning tasks across more than 20 domains and training them jointly, Intern-S2-Preview-397B achieves leading general-reasoning performance among open-source models and strong results in specialized scientific tasks such as biomolecular interaction design and material structure generation; General and scientific long-horizon agents. By connecting multiple agent frameworks to large-scale sandboxed environments for black-box agentic reinforcement learning, Intern-S2-Preview-397B improves generalization and raises the capability ceiling for long-horizon tasks in both general and scientific domains

We introduce Intern-S2-Preview-35B, an efficient 35B scientific multimodal foundation model. Beyond conventional parameter and data scaling, Intern-S2-Preview-35B explores task scaling: increasing the difficulty, diversity, and coverage of scientific tasks to further unlock model capabilities.

By extending professional scientific tasks into a full-chain training pipeline from pre-training to reinforcement learning, Intern-S2-Preview-35B achieves performance comparable to the trillion-scale Intern-S1-Pro on multiple core professional scientific tasks, while using only 35B parameters (continued pretrained from Qwen3.5). At the same time, it maintains strong general reasoning, multimodal understanding, and agent capabilities.

Features

Scientific task scaling with full-chain training. Intern-S2-Preview-35B scales hundreds of professional scientific tasks from pre-training to RL, enabling strong performance across multiple specialized domains at only 35B parameters. It further strengthens spatial modeling for small-molecule structures and introduces real-valued prediction modules, making it the first open-source model with both material crystal structure generation capability and strong general capabilities; Enhanced agent capabilities for scientific workflows. Intern-S2-Preview-35B significantly improves agentic abilities over the previous generation, achieving strong results on multiple scientific agent benchmarks; Efficient RL reasoning with MTP and CoT compression. During RL, Intern-S2-Preview-35B adopts shared-weight MTP with KL loss to reduce the mismatch between training and inference behavior, substantially improving MTP accept rate and token generation speed. It also introduces CoT compression techniques to shorten responses while preserving strong reasoning capability, achieving improvements in both performance and efficiency

We introduce Intern-S1-Pro, a trillion-scale MoE multimodal scientific reasoning model. Intern-S1-Pro scales to 1T total parameters with 512 experts, activating 8 experts per token (22B activated parameters). The model delivers top-tier performance on advanced reasoning benchmarks and achieves leading results across key AI4Science domains (chemistry, materials, life-science, earth, etc.), while maintaining strong general multimodal and text capabilities.

Features

State-of-the-art scientific reasoning, competitive with leading closed-source models across AI4Science tasks; Strong general multimodal performance on various benchmarks; Trillion-scale MoE training efficiency with STE routing (dense gradient for router training) and grouped routing for stable convergence and balanced expert parallelism; Fourier Position Encoding (FoPE) + upgraded time-series modeling for better physical signal representation; supports long, heterogeneous time-series (10^0–10^6 points)

We introduce Intern-S1, our previous-generation open-source multimodal reasoning model. Intern-S1 combines strong general-task capabilities with state-of-the-art performance on a wide range of scientific tasks, rivaling leading closed-source commercial models.

Built upon a 235B MoE language model (Qwen3) and a 6B Vision encoder (InternViT), Intern-S1 has been further pretrained on 5 trillion tokens of multimodal data, including over 2.5 trillion scientific-domain tokens. This enables the model to retain strong general capabilities while excelling in specialized scientific domains such as interpreting chemical structures, understanding protein sequences, and planning compound synthesis routes, making Intern-S1 to be a capable research assistant for real-world scientific applications.

We also released Intern-S1-mini, a lightweight version of Intern-S1, which contains an 8B language model and a 0.3B vision encoder.

Features

Strong performance across language and vision reasoning benchmarks, especially scientific tasks; Continuously pretrained on a massive 5T token dataset, with over 50% specialized scientific data, embedding deep domain expertise; Dynamic tokenizer enables native understanding of molecular formulas, protein sequences, and seismic signals

We evaluate Intern-S2-397B, Intern-S2-Preview-397B, Intern-S2-Preview-35B, and Intern-S1 series models on various benchmarks, including general and scientific datasets. We report performance comparisons with recent VLMs and LLMs below.

Intern-S2-397B

Note: Underline means the best performance among open-source models, and bold indicates the best performance among all models.

We use OpenCompass, VLMEvalKit, and AgentCompass to evaluate all models. For text reasoning benchmarks, Intern-S2-397B is evaluated with a maximum inference length of 256K tokens; for multimodal benchmarks, the maximum inference length is 64K tokens.

Intern-S2-Preview-397B

Note: Underline means the best performance among open-source models, and bold indicates the best performance among all models.

We use OpenCompass, VLMEvalKit, and AgentCompass to evaluate all models. For text reasoning benchmarks, Intern-S2-Preview-397B is evaluated with a maximum inference length of 256K tokens; for multimodal benchmarks, the maximum inference length is 64K tokens.

Intern-S2-Preview-35B

Note: Underline means the best performance among open-sourced models, Bold indicates the best performance among all models.

Intern-S1-Pro

Note: Underline means the best performance among open-sourced models, Bold indicates the best performance among all models.

Intern-S1

Note: ✅ means the best performance among open-sourced models, 👑 indicates the best performance among all models.

Intern-S1-mini

We use the OpenCompass and VLMEvalKit to evaluate all models. Please refer to this page to quickly start the text-only evaluation task.

User Guide

Intern-S2-397B, Intern-S2-Preview-397B, Intern-S2-Preview-35B, and Intern-S1 series models can be deployed using any of the following LLM inference frameworks:

LMDeploy; vLLM; SGLang

Detailed deployment examples for these frameworks are available in the following guides:

Fine-tuning

See this documentation for more details.

License

This project is released under the Apache 2.0 license.

Citation

If you find this work useful, feel free to give us a cite.

A Scientific Multimodal Foundation Model