Shanghai AI Laboratory introduces Sage S2.1 Flash, a scientific multimodal model capable of complex reasoning across molecular graphs, protein structures, and specialized software tools.

On November 15, 2026, the Shanghai AI Laboratory (InternLM team) announced the release of Sage S2.1 Flash, a groundbreaking multimodal model specifically engineered for the rigorous demands of scientific research. Unlike general-purpose LLMs that often struggle with the nuances of specialized data, Sage S2.1 Flash is designed for high-stakes tasks requiring multiple rounds of iterative analysis, verification, and tool-assisted reasoning.
This release marks a significant milestone for the Sage family, a new line of models from the InternLM team known for pushing the boundaries of open-source AI. By bridging the gap between natural language instructions and complex scientific data formats, Sage S2.1 Flash provides researchers with an intelligent agent capable of not just observing, but actively participating in the scientific discovery process.
At the core of Sage S2.1 Flash is the Sage SciPerception unified representation architecture. This architecture allows the model to encode a vast array of scientific inputs into a shared representation space. This enables the model to reason about diverse data types simultaneously with natural language instructions, facilitating complex tasks like site localization, quantitative analysis, and structure inference.
The model is not merely a passive observer; it is a highly capable agent. It integrates scientific data understanding with advanced information retrieval, coding, and tool-use capabilities. Sage S2.1 Flash can execute specialized software, monitor intermediate outputs, analyze visual feedback from simulation tools, and autonomously adjust its subsequent research steps based on those results.
Sage S2.1 Flash has demonstrated industry-leading performance across several specialized benchmarks, particularly in tasks that require high-precision reasoning and agentic behavior. Most notably, it achieved a score of 90.3 on MLE-bench Lite, surpassing the highly regarded DeepSeek-V4.1-1017, which scored 89.2.
The model also shows exceptional proficiency in administrative and data-heavy scientific workflows. It secured first place on OfficeQA Pro with a score of 72.2, Harvey LAB with 25.0, and SpreadsheetBench 2 with 37.1. These results indicate a robust ability to handle complex document structures and data manipulation tasks that are critical in laboratory environments.
As Sage S2.1 Flash is primarily released as an open-weights model, the primary cost associated with its use will be the infrastructure required for self-hosting. For developers looking to deploy at scale, the model's compatibility with vLLM and SGLang ensures optimized resource utilization and lower latency.
While the Shanghai AI Laboratory provides the weights freely, users can access the model via various hosting providers or through the official discovery platform for testing purposes.
The versatility of Sage S2.1 Flash makes it an ideal candidate for several high-impact domains. In drug discovery, the model can be used to analyze protein-ligand interactions by processing 3D protein structures and molecular graphs. In materials science, it can assist in predicting the properties of new crystal structures through iterative simulation and feedback loops.
Beyond pure science, its high performance in spreadsheet and office benchmarks makes it a powerful tool for automated data science workflows, complex RAG (Retrieval-Augmented Generation) systems involving scientific literature, and autonomous coding agents specialized in scientific computing libraries.
Developers can immediately begin working with Sage S2.1 Flash through several channels. The open weights are available on Hugging Face (non-gated) and ModelScope, allowing for easy integration into existing machine learning pipelines and local environments.
For those who wish to test the model's capabilities before deploying it locally, the Shanghai AI Laboratory provides a hands-on testing environment at discovery.intern-ai.org.cn. This platform allows researchers to interact with the model and explore its multimodal reasoning capabilities in real-time.
API Pricing β Context: Open Weights