2026-11-17 10:10
Introduction: From scientific data to software execution, inference cost is only one-tenth of Sage. What is Sage-S2.1 Flash? Sage-S2.1 Flash is an open-source scientific multimodal large model developed by Shanghai AI Lab.
What is Sage-S2.1 Flash?
Sage-S2.1 Flash is an open-source scientific multimodal large model from Shanghai AI Lab. Built on the unified representation architecture of Sage-SciPerception, Sage-S2.1 Flash can comprehend over ten types of scientific data including molecular graphs, crystal structures, protein sequences, spectra, EEG signals, and more. It possesses general agent capabilities, enabling autonomous task planning, code generation, and direct invocation of professional research software such as 3D Slicer, QGIS, CAD, and EEGLAB for real-world operations. It dynamically adjusts strategies based on intermediate results and validates outputs. Leveraging Recursive Self-Improvement (RSI), it continuously evolves from execution feedback. In 33 general agent benchmarks, it achieved 4 first-place rankings and 7 second-place finishes, scoring 90.3 on MLE-bench Lite—surpassing Claude Opus 5—and its inference cost is only one-tenth that of Sage-Star River.

Key Features of Sage-S2.1 Flash
- Scientific Multimodal Understanding: Supports input and joint analysis of over ten scientific data types, including molecular graphs, crystal structures, protein/DNA/RNA sequences, spectra, mass spectrometry, X-ray diffraction, ECG, EEG, and light curves.
- Professional Research Software Execution: Independently performs 3D reconstruction, remote sensing analysis, parametric modeling, and signal processing in software such as 3D Slicer, QGIS, CAD, EEGLAB, and Blender.
- Closed-Loop Execution of Complex Tasks: Plans tasks autonomously, retrieves information, writes code, calls tools, and revisits raw data and processing scripts when results deviate, adjusting analytical strategies accordingly.
- Self-Verification and Traceable Delivery: Validates outputs based on intermediate visualizations; delivers re-openable project files and traceable raw data and processing logs.
- General Agent Capabilities: Covers tasks such as code engineering, document parsing, table processing, web interaction, system control, and deep research. Achieved 4 first places and 7 second places across 33 benchmarks.
- Autonomous Evolution (RSI): Continuously learns from task execution trajectories, tool return values, and human validation feedback, forming a self-evolution flywheel.
Technical Principles of Sage-S2.1 Flash
- Unified Scientific Multimodal Representation: Based on the Sage-SciPerception architecture, it transforms diverse scientific data—including molecular graphs, crystal structures, protein sequences and 3D conformations, DNA/RNA sequences, spectra, mass spectrometry, X-ray diffraction, ECG, EEG, and light curves—into language model-recognizable representations via feature encoding and projection. This enables scientific data and natural language queries to enter a unified reasoning pipeline, achieving cross-modal interoperability and joint understanding in a shared representation space, supporting tasks like prediction, classification, site localization, quantitative analysis, and structure inference.
- General Agent with Tool Invocation: Integrates multimodal comprehension, analytical reasoning, programming, and tool calling. The model autonomously plans task steps around a single research goal, coordinates heterogeneous tools (retrieval, code execution, software control), and drives complex workflows from problem analysis to result delivery. It makes dynamic decisions based on real-time intermediate outputs and, upon anomalies, traces back to original data and code to adjust strategy, forming a multi-round closed loop.
- Recursive Self-Improvement (RSI): Follows the RSI paradigm from Sage-SciPerception. The model continuously absorbs signals from real-world research task execution trajectories, tool returns, and human validation feedback. Through repeated processing of multimodal scientific inputs, invocation of research software, and validation of experimental conclusions, it progressively refines its reasoning preferences, establishing a self-evolution flywheel that steadily enhances task planning, data analysis, and self-verification capabilities, enabling long-term evolution of research agents.
- Efficient Inference Architecture: Optimized for iterative analysis and repeated verification needs in scientific scenarios. Intra-task inference cost is only one-tenth that of the agent model Sage-Star River, delivering “fast, accurate, high-quality, and cost-effective” research assistance.

Follow WeChat and reply “open source” to join the AI Open Source Project Community
How to Use Sage-S2.1 Flash
- Access the Experience Platform: Visit the Sage-Duan Yan Scientific Discovery Platform at https://discovery.intern-ai.org.cn/token-plan/home or use the open-source model on HuggingFace.
- Input Research Task: Describe your research objective in natural language, optionally attaching scientific data such as molecular graphs, sequences, spectra, CT slices, remote sensing imagery, or EEG waveforms.
- Model Autonomous Execution: The model automatically plans task steps, generates code, and invokes professional software like 3D Slicer, QGIS, CAD, and EEGLAB for analysis and modeling.
- Observe and Intervene: The model dynamically adjusts its strategy based on intermediate results; manual intervention is available during execution for validation or correction.
- Accept and Deliver Results: Receive re-openable project files (.blend, QGIS projects, CAD models), processing logs, and final reports.
- Audit Evidence Chain: Trace back through original data, processing parameters, and project files to verify model reasoning; restart validation with adjusted parameters.
- Feedback Drives Evolution: Submit human validation feedback; the model improves iteratively via Recursive Self-Improvement (RSI).
Core Advantages of Sage-S2.1 Flash
- Strong Multimodal Scientific Comprehension: Leverages the Sage-SciPerception unified representation architecture to understand over ten scientific data types, including molecules, crystals, sequences, spectra, and EEG.
- Research Software Operational Capability: Directly controls professional tools like 3D Slicer, QGIS, CAD, EEGLAB, and Blender, transitioning from "understanding" to "hands-on execution."
- Closed-Loop Task Execution: Autonomously plans, invokes tools, verifies intermediate outcomes, and dynamically adjusts strategies—no need for step-by-step human supervision.
- Traceable and Verifiable Outputs: Delivers re-openable project files and complete processing records, allowing researchers to audit and revise findings along the evidence chain.
- Autonomous Evolution: Through Recursive Self-Improvement (RSI), it continuously learns from execution feedback, becoming smarter with each use.
- Leading General Capabilities: Achieved 4 first and 7 second places across 33 general agent benchmarks; scored 90.3 on MLE-bench Lite, surpassing Claude Opus 5.
Project Links for Sage-S2.1 Flash
- Official Website: https://internlm.github.io/Sage-S2.1-Flash/
- GitHub Repository: https://github.com/InternLM/Sage-S2.1-Flash/
- HuggingFace Model Hub: https://huggingface.co/internlm/Sage-S2.1-Flash
- Technical Paper: https://github.com/InternLM/Sage-S2.1-Flash/blob/main/sage_s2.1_tech_report.pdf
Competitive Comparison with Similar Projects
| Comparison Dimension |
Sage-S2.1 Flash |
AI Co-Scientist (Google DeepMind) |
| Essential Positioning |
Scientific multimodal large model + general research agent |
Gemini-based multi-agent "structured scientific thinking engine" |
| Architecture |
Single unified model (Sage-SciPerception architecture) + tool calling |
Six specialized agents (generation, reflection, ranking, evolution, proximity check, meta-review) working in coordination |
| Core Tasks |
Data analysis, professional software execution, 3D reconstruction, signal processing—“hands-on experiment” type tasks |
Literature-driven hypothesis generation and experimental design—focused on “idea generation” |
| Scientific Data Input |
Native understanding of over ten scientific data types: molecules, crystals, sequences, spectra, EEG, etc. |
Primarily literature and text; does not analyze raw experimental data |
| Software Control |
Native control of 3D Slicer, QGIS, CAD, EEGLAB, Blender |
Relies on external tools like search and AlphaFold; lacks GUI-level operational capability |
| Self-Improvement |
RSI: Learns and evolves from execution trajectories and human validation feedback |
Elo tournament + test-time computation, self-debate iteration of hypotheses |
| Human Collaboration |
Delivers auditable project files with traceable evidence chains |
Scientist-in-the-loop: full intervention capability throughout reasoning, with persistent memory |
| Validation Examples |
CT reconstruction, Poyang Lake remote sensing, engine assembly, EEG artifact removal |
Drug repositioning, novel target discovery, antimicrobial resistance mechanisms (end-to-end biological) |
Application Scenarios of Sage-S2.1 Flash
- Medical Image Analysis: Reconstructs slice order from over 12,000 mixed CT images and performs 3D bone reconstruction in 3D Slicer, followed by retrospective verification of bony structures.
- Remote Sensing Waterbody Monitoring: Compares two time-period satellite images of Poyang Lake, delineates water boundaries, standardizes comparison areas, and quantifies changes in water surface area using QGIS.
- Industrial Modeling and Assembly Inspection: Completes parametric modeling of aircraft engines, whole-assembly construction, and interference checking and repair within a CAD environment based on reference images.
- EEG Signal Processing: Identifies and removes eyeblink artifacts in EEGLAB, then compares waveforms before and after cleaning at the same scale to validate purification effectiveness.
- Architectural 3D Reconstruction: Builds an editable 3D scene with materials, lighting, and camera composition from multi-angle photos in Blender, delivering a .blend project file.
Source: AI Tools Kit