Intern S1 Pro targets AI-for-science with trillion-parameter MoE model

Intern S1 Pro is a trillion-parameter mixture-of-experts model designed specifically for scientific research, spanning chemistry, materials science, life sciences, and earth science. While the raw parameter count is huge, only about 22 billion parameters are active at inference, improving efficiency.

On domain benchmarks, Intern S1 Pro claims state-of-the-art performance versus both open and closed competitors, beating offerings like Kimi K 2.5 and Gemini 3 Pro across many science tasks. It’s optimized for advanced reasoning in technical domains rather than general chat, making it a better fit for labs and enterprises that need on-prem scientific assistants.

The downside is scale: model weights are roughly 919GB, so running Intern S1 Pro locally requires a serious GPU cluster or in-house data center. For organizations that can meet that bar, the project page lays out download links and setup instructions for deploying the model on private infrastructure.

For universities and research hubs in SEA investing in AI-for-science, Intern S1 Pro offers a path to high-end reasoning without relying solely on foreign cloud APIs.