Sage S2.1 Flash Technical Report

Kun Cai, Haodong Chen, Guangran Cheng, Qingxiu Cheng, Xu Dai, Guanting Dong, Qiming Ge, Lixin Gu, Tao Gui, Honglin Guo, Yuyang Hu, Zixian Huang, Jiajie Jin, Alexander Lam, Jiaqiang Li, Yining Li, Jiahang Lin, Yanjiang Liu, Xinyu Lu, Haijun Lv, Ningsheng Ma, Zerun Ma, Minlong Peng, Jianmin Qian, Huayu Sha, Junlin Shang, Shuo Shen, Qisheng Su, Jiajun Sun, Yanan Sun, Zhichao Sun, Ran Tao, Zhongbo Tian, Guoqiang Wang, Hui Wang, Rui Wang, Yiding Wang, Zhecan Wang, Hao Xiang, Xinchen Xie, Shuhao Xing, Xiaoyu Xing, Wanghan Xu, Xinyu Yang, Yajie Yang, Dingbo Yuan, Yuhang Zang, Hongjie Zhang, Chengfeng Zhao, Haoran Zhao, Penghao Zhao, Ruojun Zhou, Dongsheng Zhu, Yicheng Zou, Qiye Cai, Yifei Cao, Xinmeng Che, Hongji Chen, Jiabei Chen, Jiadong Chen, Jiahao Chen, Jiayi Chen, Xinjie Chen, Yujia Chen, Xin Deng, Yi Dong, Wenjie Fu, Jianing Geng, Chenyang Gu, Ding Han, Feiyang Hao, Haotan He, Qiushi He, Xufeng He, Xin Hong, Jie Hou, Binze Hu, Yongkang Hu, Zijian Hu, Enfu Huang, Junhao Huang, Changhao Jiang, Huicheng Jiang, Jiazhen Jiang, Shufan Jiang, Wenqing Jing, Jiahao Kuang, Bowen Lai, Bo Li, Qi Li, Zhuoqun Li, Tianhao Liang, Jiaxiang Liu, Junzhang Liu, Shaofan Liu, Shuainan Liu, Tong Liu, Yi Liu, Ziyu Liu, Zhonghang Lu, Jianwen Luo, Xuan Luo, Yanyi Luo, Huijie Lv, Mingyu Ma, Houcheng Min, Hongjun Mo, Songhao Ni, Chengjun Pan, Qiyuan Peng, Xiaoxuan Peng, Wanying Ren, Runhui Ruan, Jifei Shan, Zixin Shang, Bing Shao, Zhuohui Sheng, Jiayang Shi, Qingyun Shi, Yang Shu, Aierpanjiang Simayi, Sirui Song, Yuxiao Song, Jiahao Sun, Zhe Sun, Wenzhe Tan, Guoxiang Wang, Hanchen Wang, Jie Wang, Pengbo Wang, Qing Wang, Weihang Wang, Xiaotao Wang, Yuhui Wang, Zhekai Wang, Zihan Wang, Zile Wang, Jinke Wu, Liwei Wu, Xuansheng Wu, Yida Wu, Yihao Wu, Yinqi Wu, Yue Wu, Yuhan Wu, Yuwa Wu, Zhiheng Xi, Ying Xiao, Zhide Xie, Caijun Xu, Xingle Xu, Yichen Xu, Yongfeng Xu, Xiaolei Yang, Zhixiong Yang, Jiacheng Yao, Dingcheng Yi, Shihong Yi, Jinhui Yin, Peizhou Yu, Shan Yu, Hao Yuan, Junjie Yuan, Xiongyao Zha, Bo Zhang, Caixian Zhang, Kaicheng Zhang, Pengxiang Zhang, Qiuyinzhe Zhang, Wei Zhang, Xiaoqian Zhang, Xinsen Zhang, Jiyuan Zhao, Shuyu Zhao, Pujun Zheng, Zheyu Zheng, Xiaoxue Zhong, Peiheng Zhou, Xiaohao Zhou, Xinyu Zhou, Fangyi Zhu, Guanru Zhu, Yulun Zhu, Jiabao Zhuang, Pengfei Cao, Zhicheng Dou, Xianpei Han, Ben He, Guoxiu He, Xuanjing Huang, Tao Ji, Yu-Gang Jiang, Dahua Lin, Hongyu Lin, Kang Liu, Yaojie Lu, Le Sun, Ji-Rong Wen, Jun Zhao, Yutao Zhu, Bowen ZhouarXiv 2026 · arXiv

TL;DR

Sage S2.1 Flash is a 320-billion-parameter Mixture-of-Experts (MoE) model designed to bridge the gap between scientific reasoning and autonomous agentic action. By integrating specialized encoders for molecular and biological data with a "ToolSpace" framework for verifiable task execution, it achieves state-of-the-art performance on benchmarks ranging from machine learning engineering to complex professional workflows.

What Problem Does This Paper Address?

While general-purpose LLMs have become remarkably capable, they often stumble in two specialized areas:

  1. Scientific Perception: Most models treat scientific data (like molecular graphs, protein sequences, or chemical spectra) as mere text or images, losing the underlying structural and physical nuances.
  2. Reliable Agentic Execution: Performing long-horizon tasks in interactive environments—such as coding, searching, or using specialized software—is prone to "hallucinated" actions. Current models often lack a robust way to verify if their actions actually achieved the intended goal in a real-world environment.

The Sage S2.1 Flash team tackles these by creating a model that "perceives" science natively and "acts" within verifiable, execution-based environments.

Methods

The authors introduce several novel architectural and training components:

  • Sage Sci-Perception: Instead of relying solely on a standard vision or text encoder, the model uses dedicated, frozen scientific encoders (e.g., MoleculeSTM for graphs, ESM-2 for proteins). These representations are mapped into a shared backbone via modality-specific projectors. For numerical data like spectra, they use "numerical patching" to maintain measurement precision. Overview of the Sage Sci-Perception interface
  • 320B MoE Backbone: The core is a massive Mixture-of-Experts model that routes tokens through specialized subsets of parameters, allowing for high capacity with manageable computational costs.
  • ToolSpace Framework: To train the model's agentic skills, the authors developed ToolSpace, which reconstructs a "tool-composition space" from real execution traces. This allows them to synthesize complex, multi-step tasks across different environments (e.g., GitHub, SQL, Slack) that are grounded in real-world logic. Overview of ToolSpace for cross-environment data synthesis
  • Verifiable Post-Training: The model is trained using Group Relative Policy Optimization (GRPO). Crucially, rewards are not based on "looking correct" but on verifiable outcomes: Did the code pass the unit tests? Did the environment state change as expected? Did the submitted artifact meet the requirements?
Overview of the Sage Sci-Perception interface
Overview of ToolSpace for cross-environment data synthesis

Datasets

The model was evaluated on a massive suite of 33 benchmarks across seven domains:

  • Agentic Coding: Terminal-Bench 2.1, DeepSWE, SWE-bench Pro.
  • Productivity: SpreadsheetBench 2, ALE-CLI, GDPval.
  • Tool Use: OfficeQA Pro, Harvey LAB, AutomationBench.
  • ML Engineering: MLE-bench Lite, PostTrainBench, PaperBench.
  • Vision & Search: OSWorld, WebArena, DeepSearchQA.

The authors also utilized the MCP (Model Context Protocol) ecosystem to collect real-world tool interaction trajectories for training.

Key Insights and Ideas

  • Science as a Native Language: By using specialized encoders, the model doesn't just "see" a picture of a molecule; it understands the atomic connectivity. This allows it to perform tasks like structure editing and property prediction far more accurately than generalist models.
  • The "Verifiability" Pivot: The shift from training on static text to training in "verifiable environments" is a major theme. If a model's action can be checked by a compiler or a database query, the feedback loop for reinforcement learning becomes significantly more reliable.
  • Visual Feedback in Computer Use (CUA): For desktop and browser tasks, the model is trained to interpret visual feedback from its own actions. It doesn't just click a button; it looks at the screen to see if the button click worked before deciding the next step. Overview of the CUA training data construction pipeline
  • Iterative Trajectory Analysis: The authors use a "Human-guided iterative trajectory analysis" (Figure 6) to debug model failures, finding that Sage S2.1 Flash is particularly good at resolving uncertainties by running "experimental checks"—essentially trying different configurations until one works.
Overview of the CUA training data construction pipeline

Key Results

Sage S2.1 Flash consistently ranks at the top of its class, particularly in engineering and research tasks:

  • 1st Place: OfficeQA Pro (72.2), MLE-bench Lite (90.3 HumanRank), Harvey LAB, and SpreadsheetBench 2.
  • 2nd Place: OSWorld-Verified (85.2), WebArena-Verified (67.6), and DeepResearch Bench II (54.2).
  • ML Engineering: It leads primary baselines on PostTrainBench (39.2) and PaperBench Code-Dev (88.0), demonstrating a high capacity for automating actual research workflows. Results on 15 selected benchmarks
Results on 15 selected benchmarks

Limitations and Future Work

Despite its strengths, the authors identify several areas for growth:

  • Repository-Level Coding: While excellent at "patch-level" fixes, there is still room to improve on massive, multi-file repository-level generation.
  • Web Search: While competitive, the integration of multi-step search and information synthesis can be further refined.
  • End-to-End Scientific Discovery: The next step is moving from individual benchmarks to "complete scientific workflows"—where the model autonomously selects evidence, conducts experiments, and produces a final, scientifically correct report.

Summary updated

Abstract

We introduce Sage S2.1 Flash, a 320-billion-parameter general-purpose mixture-of-experts model designed to reason over scientific evidence and act in interactive environments. Sage Sci-Perception uses dedicated encoders for molecular graphs and biological sequences and represents spectra as ordered patches of measurements. Modality-specific projectors map these representations into the input space of a shared backbone, which processes them alongside text and images. To develop the model's agent capabilities, we construct verifiable environments for tool use, search, and coding, as well as specialized workflows involving machine learning experiments and scientific software. Task completion is checked against execution results, changes in environment state, or submitted artifacts. We evaluate Sage S2.1 Flash on 33 benchmarks across seven domains. In the primary comparison against five baselines, it ranks first on OfficeQA Pro (72.2), MLE-bench Lite (90.3 HumanRank), Harvey LAB, and SpreadsheetBench 2, and second on OSWorld-Verified (85.2), WebArena-Verified (67.6), and DeepResearch Bench II (54.2). These results show particular strengths in using documentary evidence, producing verifiable workbooks, and carrying out machine learning tasks, alongside competitive performance in desktop and browser workflows that require the model to act on visual feedback.
BibTeX
@misc{cai2026sage,
  author = {Kun Cai and Haodong Chen and Guangran Cheng and Qingxiu Cheng and Xu Dai and Guanting Dong and Qiming Ge and Lixin Gu and Tao Gui and Honglin Guo and Yuyang Hu and Zixian Huang and Jiajie Jin and Alexander Lam and Jiaqiang Li and Yining Li and Jiahang Lin and Yanjiang Liu and Xinyu Lu and Haijun Lv and Ningsheng Ma and Zerun Ma and Minlong Peng and Jianmin Qian and Huayu Sha and Junlin Shang and Shuo Shen and Qisheng Su and Jiajun Sun and Yanan Sun and Zhichao Sun and Ran Tao and Zhongbo Tian and Guoqiang Wang and Hui Wang and Rui Wang and Yiding Wang and Zhecan Wang and Hao Xiang and Xinchen Xie and Shuhao Xing and Xiaoyu Xing and Wanghan Xu and Xinyu Yang and Yajie Yang and Dingbo Yuan and Yuhang Zang and Hongjie Zhang and Chengfeng Zhao and Haoran Zhao and Penghao Zhao and Ruojun Zhou and Dongsheng Zhu and Yicheng Zou and Qiye Cai and Yifei Cao and Xinmeng Che and Hongji Chen and Jiabei Chen and Jiadong Chen and Jiahao Chen and Jiayi Chen and Xinjie Chen and Yujia Chen and Xin Deng and Yi Dong and Yichen Dong and Wenjie Fu and Jianing Geng and Chenyang Gu and Ding Han and Han Han and Feiyang Hao and Haotan He and Qiushi He and Xufeng He and Xin Hong and Jie Hou and Binze Hu and Yongkang Hu and Zijian Hu and Enfu Huang and Junhao Huang and Changhao Jiang and Huicheng Jiang and Jiazhen Jiang and Shufan Jiang and Wenqing Jing and Jiahao Kuang and Bowen Lai and Bo Li and Qi Li and Qilong Li and Zhuoqun Li and Tianhao Liang and Jiaxiang Liu and Junzhang Liu and Shaofan Liu and Shuainan Liu and Tong Liu and Yi Liu and Ziyu Liu and Zhonghang Lu and Jianwen Luo and Xuan Luo and Yanyi Luo and Huijie Lv and Mingyu Ma and Houcheng Min and Hongjun Mo and Songhao Ni and Chengjun Pan and Qiyuan Peng and Xiaoxuan Peng and Wanying Ren and Runhui Ruan and Jifei Shan and Zixin Shang and Bing Shao and Zhuohui Sheng and Jiayang Shi and Qingyun Shi and Yang Shu and Aierpanjiang Simayi and Sirui Song and Yuxiao Song and Jiahao Sun and Zhe Sun and Wenzhe Tan and Guoxiang Wang and Hanchen Wang and Jie Wang and Pengbo Wang and Qing Wang and Rui Wang and Weihang Wang and Xiaotao Wang and Yuhui Wang and Zhekai Wang and Zihan Wang and Zile Wang and Jinke Wu and Liwei Wu and Xuansheng Wu and Yida Wu and Yihao Wu and Yinqi Wu and Yue Wu and Yuhan Wu and Yuwa Wu and Zhiheng Xi and Ying Xiao and Zhide Xie and Caijun Xu and Xingle Xu and Yichen Xu and Yongfeng Xu and Xiaolei Yang and Zhixiong Yang and Jiacheng Yao and Dingcheng Yi and Shihong Yi and Jinhui Yin and Peizhou Yu and Shan Yu and Hao Yuan and Junjie Yuan and Xiongyao Zha and Bo Zhang and Caixian Zhang and Kaicheng Zhang and Pengxiang Zhang and Qiuyinzhe Zhang and Wei Zhang and Xiaoqian Zhang and Xinsen Zhang and Jiyuan Zhao and Shuyu Zhao and Pujun Zheng and Zheyu Zheng and Xiaoxue Zhong and Peiheng Zhou and Xiaohao Zhou and Xinyu Zhou and Fangyi Zhu and Guanru Zhu and Yulun Zhu and Jiabao Zhuang and Pengfei Cao and Zhicheng Dou and Xianpei Han and Ben He and Guoxiu He and Xuanjing Huang and Tao Ji and Yu-Gang Jiang and Dahua Lin and Hongyu Lin and Kang Liu and Yaojie Lu and Le Sun and Ji-Rong Wen and Jun Zhao and Yutao Zhu and Qi Zhang and Bowen Zhou},
  title = {{Sage S2.1 Flash Technical Report}},
  year = {2026},
  eprint = {2611.21072},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.21072},
}
Figure