Technology & AIAnalysis
Shanghai AI Lab Releases Open-Source Shusheng-S2 for Scientific Research
The multimodal model aims to handle long-cycle scientific tasks alongside domestic hardware and national research facilities.
Share
Photo by Pavel Danilyuk on Pexels
The Brief
The Shanghai Artificial Intelligence Laboratory has unveiled Shusheng-S2, an open-source multimodal foundation model designed for long-cycle scientific research and autonomous agent tasks. Announced at the 2026 Pujiang Innovation Forum, the model demonstrates specialized performance in life sciences and materials design while being co-optimized with the domestic Ascend computing ecosystem. Concurrently, the laboratory upgraded its Duanyan scientific discovery platform, integrating Shusheng-S2 and a physical world model with major national laboratory apparatus to establish closed-loop, automated experimental research pipelines.
Why it matters
Sustained, long-horizon exploration has long been a core bottleneck for general-purpose large models in natural sciences. By combining model architecture with physical automation and national research infrastructure, Shusheng-S2 and its companion platform seek to bridge computational simulation with physical experimentation, advancing artificial intelligence for science from isolated calculations into verifiable, end-to-end laboratory workflows.
China context
Faced with ongoing constraints on external high-performance hardware, Chinese research institutes are deliberately optimizing major foundational models for domestic computing ecosystems such as Huawei's Ascend. Aligning these models directly with state-level research facilities—including synchrotron light sources and neutron facilities—reflects Beijing's strategy to harness institutional scientific infrastructure to build self-reliant AI for science capabilities.
Editor's View
EDITOR'S VIEW — Analysis and inference, not factual reporting.
The release of Shusheng-S2 highlights a deliberate transition in AI for science from purely digital reasoning toward direct physical grounding. While benchmark improvements on domain tasks like molecular design are noteworthy, the real operational test lies in whether automated platforms can consistently generate reproducible, peer-validated experimental results when scaled beyond pilot setups.
What to watch
- Public availability and licensing terms of the Shusheng-S2 code and model weights repository
- Peer-reviewed research and experimental outcomes originating from closed-loop automated workflows in protein and materials synthesis
- Performance and deployment metrics of Ascend hardware clusters managing high-throughput scientific data
Key Takeaways
- 1Shanghai AI Laboratory released Shusheng-S2, an open-source multimodal foundation model engineered for long-cycle scientific discovery and agent-driven research tasks.
- 2The model is co-optimized with the Ascend computing ecosystem across computation, communication, and memory management.
- 3The upgraded Shusheng Duanyan platform integrates Shusheng-S2 and physical world model W0 to direct physical laboratory equipment and national research facilities.
- 4Targeted research domains include life sciences, materials, semiconductors, nuclear fusion, quantum technology, and atmospheric science.
The Shanghai Artificial Intelligence Laboratory unveiled Shusheng-S2, an open-source multimodal foundation model tailored for long-cycle scientific workflows, during the 2026 Pujiang Innovation Forum, according to People's Daily. The model is designed to support prolonged research timelines by strengthening autonomous agent capabilities, scientific knowledge expansion, and complex reasoning.
According to the laboratory, Shusheng-S2 achieves competitive general-purpose capabilities across reasoning, code, and agent evaluations while outperforming comparable open-source and proprietary models on domain-specific benchmarks such as Biology-Instructions and Mol-Instructions. The laboratory also confirmed that Shusheng-S2 has undergone co-optimization with the domestic Ascend computing ecosystem, targeting computational throughput, inter-chip communication, and memory efficiency to establish an integrated software-hardware foundation.
Alongside the model, the Shanghai AI Lab released an upgraded version of its Shusheng Duanyan scientific discovery platform, making its full-stack toolset publicly accessible. The platform integrates Shusheng-S2 with Shusheng W0, a specialized physical world model designed to handle force-touch feedback and physical interaction. Together, the systems are intended to connect digital computational predictions with real-world laboratory equipment.
To bridge dry-lab computation with wet-lab execution, the laboratory announced partnerships linking Duanyan to major scientific facilities, including the Shanghai Synchrotron Radiation Facility, the China Spallation Neutron Source, an atomic manufacturing installation, and an electrosynthesis research institute. These collaborations allow automated agents to leverage precise experimental instruments for probing reaction mechanisms and scaling validation tests.
The Duanyan system currently operates across six research domains: life sciences, functional materials, semiconductors, nuclear fusion, quantum science, and atmospheric modeling. In fields including protein engineering and advanced materials, the platform has reportedly completed end-to-end dry-and-wet experimental cycles, taking candidate discoveries from initial computational hypotheses through automated physical verification.
Sources
- 服务长周期科研任务 开源基础大模型“书生-S2”发布--经济·科技--人民网 — People's Daily · 9/14/2026