Shanghai AI Lab releases Intern-S2-397B science multimodal model

Shanghai AI Laboratory has published open weights for Intern-S2-397B, a science-oriented multimodal foundation model with roughly 403 billion parameters, on Hugging Face and ModelScope. It features a pluggable Memory Decoder, scientific RL across 20+ fields and is co-optimized with Huawei's Ascend stack.

• About 403 billion parameters, open weights on Hugging Face and ModelScope
• Memory Decoder lifts Biology-Instructions from 56.92 to 60.32 without retraining the base
• Trained on scientific literature, RL across 20+ fields and sandboxed agent tasks
• Served via LMDeploy, vLLM, SGLang and an official Intern API