Paper summary: Intern-S1-Pro: Scientific Multimodal Foundation Model at Trillion Scale

Problem

General LLMs are broad but shallow on specialised science. Specialised models are deep but narrow.

Method

Intern-S1-Pro scales to one trillion parameters. Development follows the SAGE framework (Synergistic Architecture for Generalizable Experts) with Foundation, Fusion and Evolution layers (Figure 1). The XTuner and LMDeploy infrastructure enables efficient RL training at the 1T scale, while keeping precision consistent between training and inference.

Key results

From the abstract:

• stronger reasoning, image–text understanding and agent capabilities
• expertise across over 100 specialised tasks in chemistry, materials, life sciences and earth sciences
• top tier among open-source models on general capabilities, and ahead of proprietary models on the depth of specialised scientific tasks

Limitations

There’s no dedicated limitations section in the extracted text. Comparisons are the developers’ own.