Frontier Models
Intern-S2-397B: Shanghai AI Laboratory Releases 397B Multimodal Model
The InternLM team open-sources a large-scale model that scales pre-training, reinforcement learning across scientific domains, and agent environments to support long-horizon tasks.
Intern-S2-397B is a multimodal foundation model from Shanghai AI Laboratory designed for scientific reasoning and long-horizon agentic tasks.
The InternLM team at Shanghai AI Laboratory has released Intern-S2-397B as its most capable multimodal foundation model to date. The release follows the preview version labeled Intern-S2-Preview-397B and focuses on capabilities for scientific applications.
What background led to the Intern-S2-397B announcement?
Development of the model draws from earlier work in the Intern-series large models. The GitHub repository for InternLM details how the new model advances scientific multimodal capabilities through targeted scaling.
It targets long-horizon tasks that require sustained reasoning across complex scientific problems. This builds on previous releases in the InternLM lineup by expanding task coverage and agent interaction environments.
How does the vision-language pre-training paradigm operate in Intern-S2-397B?
The model employs a novel vision-language pre-training paradigm. It learns directly from raw rendered pages of scientific literature. This method jointly models symbolic semantics and visual relationships without intermediate parsing steps.
By processing the pages as they appear, the model captures both textual and visual elements in their natural context. Such direct learning reduces potential information loss that occurs in parsed pipelines.
What post-training approaches support long-horizon generalization?
Post-training consists of large-scale multi-task reinforcement learning combined with black-box agentic reinforcement learning. These are conducted in large-scale sandboxed environments to enhance performance on extended tasks.
The combination allows the model to improve generalization across diverse scientific scenarios. Task coverage spans more than 20 domains to prepare agents for varied research workflows.
- Pre-training on rendered scientific pages
- Large-scale multi-task RL
- Black-box agentic RL in sandboxed environments
- Integration of memory decoder for specialization
What improvements does the Intern-MemDec-4B provide?
The Intern-MemDec-4B serves as a memory decoder extension. It attaches to the frozen 397B backbone to enable rapid domain specialization without requiring full retraining.
Where can developers access Intern-S2-397B and related models?
The preview version and associated models are hosted on multiple platforms for broad accessibility. Users can obtain weights and documentation through established AI model repositories.
| Platform | Model Identifier | Description |
|---|---|---|
| Hugging Face | internlm/Intern-S2-Preview-397B | Primary hosting for the preview model |
| ModelScope | Shanghai_AI_Laboratory/Intern-S2-Preview-397B | Organization page listing the models |
| GitHub | InternLM/Intern-S1 | Repository with model introduction and details |
What market and stakeholder implications arise from this release?
The open-source nature of the model allows researchers and developers to build upon it for scientific applications. This could foster innovation in agentic AI systems tailored to science.
Stakeholders in the AI community gain access to a large-scale model without proprietary restrictions. It may influence how multimodal models are developed for specialized domains by demonstrating effective scaling strategies.
What reactions have experts and the community expressed?
The official announcement emphasized the model's focus on scientific research and agents. Community discussion centers on the potential for agentic workflows in research settings.
🚀 Introducing Intern-S2-397B, our new multimodal foundation model designed for long-horizon scientific research and powerful scientific agents.Intern Large Models, Official account for Intern-series large models by Shanghai AI Laboratory
What developments are anticipated next for this model series?
Continued scaling and application in additional scientific domains are likely as adoption grows. The modular approach with the memory decoder may lead to more efficient fine-tuning methods for other backbones.
Researchers may explore integration with existing agent frameworks to test long-horizon performance in real-world scientific simulations.
Frequently asked
What is the primary target of Intern-S2-397B?
It targets scientific intelligence and long-horizon agentic tasks through multimodal understanding.
How many scientific domains receive reinforcement-learning task coverage?
The model provides coverage across more than 20 scientific domains.
What is the function of the Intern-MemDec-4B component?
It acts as a memory decoder extension that attaches to the frozen backbone for domain specialization without altering the main model.
Sources
- X — Official announcement introducing the Intern-S2-397B model.
- GitHub — Introduction and scaling details for Intern-S2-397B.
- arXiv — Presentation of Intern-S2-Preview and the Biology-Instructions score improvement statistic.
- Hugging Face — Hosting and introduction of the Intern-S2-Preview-397B model.
- ModelScope — Models including Intern-S2-Preview-397B and Intern-S2-Preview-397B-FP8 listed under Shanghai_AI_Laboratory.
- KBlip — InternLM ships Intern-S2-397B, a multimodal foundation model targeting scientific reasoning and long-horizon agentic tasks.