Sage S2.1 Flash: 320B MoE Agent Tops 4 Benchmarks with Scientific Multimodal Perception
Shanghai AI Lab's Sage S2.1 Flash, a 320B MoE model, integrates scientific multimodal perception for molecules, proteins, DNA, RNA, and spectra. Trained via ToolSpace, ML Coding, and visual agent environments, it achieves four first-place and four second-place results across 33 benchmarks, demonstrating end-to-end scientific task execution in engineering, data analysis, and research.
Overview
Shanghai AI Laboratory, together with Fudan University, Chinese Academy of Sciences institutes, Renmin University, East China Normal University, Harbin Institute of Technology, Peking University, and others, released the Sage S2.1 Flash technical report. The report fully discloses the training pipeline and technical details of a 320B-parameter Mixture-of-Experts (MoE) model that unifies scientific multimodal perception, tool use, code execution, and professional software operation in a single agent.
Unified Scientific Multimodal Architecture
Sage S2.1 Flash adopts a 320B MoE architecture with dynamic expert routing to control inference compute, combined with sparse and linear attention for long-context processing. The shared backbone processes text, images, and scientific data representations in a common context, providing the reasoning foundation for complex multi-step tasks.
Sage Sci-Perception: Integrating Structured Scientific Data
To handle molecules, proteins, nucleic acids, and spectral data, the team designed the Sage Sci-Perception module. It employs specialized encoders: MoleculeSTM for molecular structure ESM-2 for protein sequences DNABERT-2 for DNA RNA-FM for RNA DiffSpectra for spectral signals
These encoders extract atomic connectivity, sequence dependencies, and continuous measurement features. Trainable modality projectors map them into the language model's unified representation space. Training combines supervised learning and reinforcement learning on verifiable tasks such as molecular property classification, protein interaction site prediction, atom counting, and structure relationship queries, using classification accuracy, set overlap, and numerical error as reward signals.
Scalable Verifiable Training Environments
ToolSpace: Automatic Construction of Complex Tool-Use Trajectories
For tool-calling capability, the ToolSpace framework extracts tool definitions, call parameters, and execution results from MCP tool services and real agent trajectories, building a tool-composition graph that captures data flow, environment state changes, and failure-recovery relationships. A state-aware random walk automatically generates cross-tool, cross-environment call sequences, with execution-condition checks, failure rollback, and resampling to ensure executability. Successful trajectories are converted into natural-language tasks with verification programs for final states, key intermediate states, and output fields. The resulting space covers 2,042 tools, 77,760 call-pattern nodes, and 804 environments . Supervised fine-tuning trajectories average 18.86 tool calls (max 209) and span 2.29 environments on average.
ML Coding: Executable, Verifiable Machine-Learning Research Tasks
An automated pipeline collects 4,936 Kaggle data sources and 2,960 research papers , plus code and data from Hugging Face and GitHub. It matches datasets with algorithm implementations via data interfaces, training entry points, and evaluation metrics, and validates environment usability in fixed-dependency containers. Tasks cover predictive modeling, algorithm implementation, paper reproduction, and model post-training. For algorithm tasks, core code (loss functions, feature transforms, parameter updates) is removed, requiring the model to complete it. Open-ended experiments require the agent to design and execute training. All tasks are packaged with Harbor into reproducible environments isolating instructions, execution environment, and private evaluation programs, verifying completion via code execution results, prediction metrics, and model weights.
Visual Agent: From Interface Perception to Continuous Operation
A two-phase strategy trains the model in dynamic visual environments:
Phase 1: Image recognition and single-step click tasks teach interface element understanding, target localization, and action selection from the current screenshot.
Phase 2: Extends to continuous operation sequences. At each step the model reads the current interface, judges progress, executes an action, and decides the next action from the new screenshot. Longer sequences and stronger step dependencies improve environment-state tracking and long-horizon planning.
Other Verifiable Environments
Additional environments cover information retrieval (multi-hop), software engineering (code development), cybersecurity (vulnerability analysis), and productivity tools (PPT, Excel, legal documents, 3D engineering design). Each domain builds a complete pipeline from data collection and task synthesis to environment construction and result verification, incorporating fact/evidence checking, program execution, environment-state inspection, and final-deliverable validation.
Benchmark Evaluation: 33 Benchmarks Across Seven Domains
The team evaluated Sage S2.1 Flash on agent programming, office productivity, tool use, ML engineering, deep research, information retrieval, and visual interaction.
Four First-Place Results
MLE-bench Lite: 90.3 HumanRank (1st among compared models)
OfficeQA Pro: 72.2 (1st)
Harvey LAB: 25.0 (1st)
SpreadsheetBench 2: 37.1 (1st)
Four Second-Place Results
PostTrainBench: 39.2 (2nd)
PaperBench: 88.0 (2nd)
OSWorld-Verified (desktop): 85.2 (2nd)
WebArena-Verified (web): 67.6 (2nd)
Real-World Scientific Applications
Engineering Modeling with Blender and CAD
Given images and design requirements, the model operates Blender and CAD software to construct 3D models of buildings, turbofan engines, and truck chassis, producing editable and inspectable engineering files.
Scientific Software Analysis
The model uses professional tools such as Fiji for microscopy image processing, EEGLAB for EEG signal analysis, and QGIS for remote-sensing data comparison.
Deep Research: Optical Quantum Memory
For a literature-driven research case on optical quantum storage, the model performs literature retrieval, experimental parameter extraction, simulation computation, visualization, and generates an interactive research application.
Resources
Paper: https://arxiv.org/abs/2611.21072
Project page: https://internlm.github.io/Sage-S2.1-Flash/
Hugging Face: https://huggingface.co/papers/2611.21072
GitHub: https://github.com/InternLM/Sage-S2.1-Flash
Demo: https://discovery.intern-ai.org.cn/token-plan/home
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
