A new open model at the intersection of science and AI agents
Designed for repeated analysis and verification, Sage S2.1 Flash combines scientific data, coding, and tools in a model with open weights.
Molecules, crystals, and proteins are among the data types Sage S2.1 Flash is designed to work with during research. Introduced by the Shanghai Artificial Intelligence Laboratory, this multimodal model combines scientific data understanding with information retrieval, programming, and specialized software use. The official overview describes a multistep workflow: run an analysis, examine intermediate results and visual feedback, then adjust subsequent steps.
More than ten scientific input types are supported, including molecular graphs, crystal structures, protein sequences, and three-dimensional protein structures. These inputs are processed through Sage SciPerception: specialized encoders extract features from each data type, and trained components convert them into representations the language model can use. Written instructions and scientific data can then be processed together. The laboratory targets tasks such as property prediction, classification, site localization, and structure inference.
Integrating these inputs requires specific preparation. The projection components are included in the model repository, but scientific encoders, preprocessing dependencies, and task-specific output modules must be installed separately. On the application side, data features must be extracted, transformed, and combined with text representations before a request is submitted. The documentation distinguishes this scientific workflow from deployment for general coding, office productivity, and computer-use tasks.
The release includes two open weights versions, with the repository specifying an MIT license for both code and weights. The MoE architecture, based on GLM-5.3-Flash, has 320 billion total parameters, with 18 billion activated during processing. The BF16 weights occupy approximately 645 GB; the vLLM and SGLang examples in the deployment guide use eight H200 GPUs. An FP8 version reduces that footprint to approximately 330 GB, with examples using four H200s. These are reference configurations rather than stated minimum hardware requirements.
The published evaluations also cover a broad range of professional tasks. In the team’s comparison, the model outperforms the five other systems on MLE-bench Lite with a score of 90.3, OfficeQA Pro with 72.2, Harvey LAB with 25.0, and SpreadsheetBench 2 with 37.1. These first-place results apply to the models and configurations listed in the table and do not establish overall superiority. On PaperBench, for example, its score of 88.0 trails several of the competitors shown. The reported results offer more evidence about digital task execution than about the validation of scientific discoveries.
Integration with inference software also requires some practical precautions. The documentation provides instructions for LMDeploy, vLLM, and SGLang, requiring the repository’s custom tokenizer to prevent certain passages in code or logs from being interpreted as internal control instructions. For vLLM, it also requires a patch: when multiple tools return results in a different order from their calls, images can be associated with the wrong results. The proposed fix in the vLLM project remains open. The guide also identifies a separate issue affecting non-streaming responses, where certain markers written as ordinary text can prematurely end reasoning or hide a tool call.