Developers introduced the multimodal model Intern-S2-397B

The team of developers has released Intern-S2-397B — a multimodal foundational model focused on scientific computing and long-term agent operation. The system contains 397 billion parameters and uses an updated pre-training paradigm to process visual and text data.

The model demonstrates high efficiency in solving complex scientific tasks. During testing, the system scored 46 points, taking 2 hours to perform operations. The architecture supports features for calling external tools, allowing the model to be integrated into automated workflows.

For deployment, Intern-S2-397B is compatible with popular inference frameworks, including LMDeploy, vLLM, and SGLang. This provides flexibility when launching the model on various computing powers, optimizing resource usage when working with large weights.

The novelty expands the capabilities for researchers working in the field of scientific intelligence. The use of specialized training paradigms allows the model to better handle narrow-profile tasks that require in-depth data analysis and sequential action planning within agent systems.