Intern-S2-397B (Shusheng-S2)

Overview

Intern-S2-397B, also called Shusheng-S2, is a scientific multimodal large model from Shanghai AI Laboratory, announced on September 13, 2026 at the Pujiang Innovation Forum and formally released as a full open-source build on Hugging Face on September 14. It is positioned as a foundation for scientific intelligence and long-horizon agents, with 397B parameters, general ability in the top tier of open models, and standout results on life-science and materials tasks covering understanding, generation and design.

Its most distinctive piece of architecture is a pluggable Memory Decoder: domain knowledge is learned by independent memory modules and attached to the base model without modifying its parameters, and at answer time the model dynamically balances contributions from the base model and the specialist memory. The official biology result: after attaching Intern-MemDec-4B, the Biology-Instructions average rises from 56.92 to 60.32, while general knowledge, reasoning and multimodal ability stay close to the base model's level.

Key Features

• Pluggable specialist memory: The Memory Decoder attaches domain knowledge as independent memory modules without touching base model parameters, so a new direction only requires training the matching memory module
• Long-horizon scientific reasoning: Leads existing open models on benchmarks that stress long, rigorous reasoning such as IMO-Proof and AdvancedMathBench, with the vendor claiming parity with top closed models
• Standout scientific results: Leads the open and closed flagship models in the same comparison table on life-science and materials tasks including Biology-Instructions, Mol-Instructions and MolecularIQ
• Multimodal scientific understanding: Learns directly from raw pages of scientific literature through visual pretraining, and jointly trains on scientific reinforcement learning tasks spanning more than 20 domains
• Black-box agentic reinforcement learning: Trains inside large-scale sandbox environments with black-box agentic RL, sharpening task decomposition, tool selection and multi-tool coordination
• Optimized for the Ascend ecosystem: Deep co-optimization with the Ascend compute ecosystem across compute, communication and memory, exploring how model capability and domestic compute infrastructure can evolve together

Use Cases

• Literature understanding, molecular design and structure generation in life science and materials research
• Competition and advanced mathematics problems that require hours of continuous reasoning
• Surveying, hypothesis generation and experiment planning inside a research workflow
• Serving as an agent foundation for multi-step scientific tasks, for example generating a 3D CAD model from a single sentence
• Plugging into the Shusheng Duanyan scientific discovery platform to connect with instruments and a real research loop

Pros

• The Memory Decoder extends domain knowledge without retraining the whole model, preserving existing general ability
• Scientific task results sit near the top of the open and closed models in its comparison table, making it usable in specialist settings
• General ability also reaches the top tier of open models, so it is not a narrow specialist
• The full 397B weights are open source for self-hosting, with an official API available as an alternative
• Co-optimized with the domestic compute ecosystem, which eases deployment for research institutions in China

Pricing

Weights are open source and can be self-hosted; the vendor offers every user a free API quota with higher limits available by application. Intern-S2-Preview-397B retires on October 31, 2026, so calls should be moved to the release build in time.

Summary

Intern-S2-397B targets research and long-horizon reasoning: teams doing life-science or materials computation, needing a model that reasons for hours on hard problems, or wiring a model into a real experiment loop should evaluate it seriously. At 397B the bar for local deployment is very high, so most users will go through the official API or a hosting platform. For routine Q&A and writing this size is not economical and a general model is the better fit.

Version History

• Intern-S2-397B ships as a full open-source release with a pluggable Memory Decoder (2026-09-14): Shanghai AI Laboratory's Shusheng-S2 release build, Intern-S2-397B, went live as open source on Hugging Face. The 397B model targets scientific intelligence and long-horizon agent ability, learns directly from raw pages of scientific literature through visual pretraining, jointly trains on scientific reinforcement learning tasks across more than 20 domains, and trains inside large-scale sandbox environments with black-box agentic RL. Architecturally it introduces a pluggable Memory Decoder that attaches domain knowledge without modifying base model parameters; with Intern-MemDec-4B attached, the Biology-Instructions average rises from 56.92 to 60.32. Official evaluations report 87.00 on FrontierScience-Olympiad, 93.56 on HMMT-2026 and 68.54 on SWE-bench-Pro, leading existing open models on IMO-Proof and AdvancedMathBench. The model is deeply co-optimized with the Ascend compute ecosystem and will plug into the Shusheng Duanyan scientific discovery platform; Intern-S2-Preview-397B retires on October 31, 2026.
• Shanghai AI Laboratory announces Shusheng-S2 with scientific ability on par with closed flagships (2026-09-13): At the AI for Science session of the 2026 Pujiang Innovation Forum, Shanghai AI Laboratory announced the Shusheng-S2 multimodal large model. The vendor says its general ability reaches the top tier of open models and that it leads existing open and closed flagship models on life-science and materials tasks such as Biology-Instructions and Mol-Instructions, and previewed NCP-ArchPreview, a latent-space language model pretrained at the trillion-token scale. The same event upgraded and fully opened the Shusheng Duanyan scientific discovery platform's full stack.