Sage S2.1 Flash: New 320B-Parameter AI Model for Scientific Tasks
The Shanghai Artificial Intelligence Laboratory has introduced Sage S2.1 Flash, a groundbreaking 320-billion-parameter mixture-of-experts (MoE) model designed for general-purpose and scientific tasks. The model was detailed in a technical report published on arXiv on November 16, 2026, and has since gained traction in the AI research community, receiving 83 upvotes on HuggingFace Daily Papers.
Advanced Multimodal Capabilities
Sage S2.1 Flash leverages a unified representation architecture called Sage SciPerception, which integrates more than ten types of scientific input, including molecular graphs, biological sequences, crystal structures, and 3D protein structures. These inputs are encoded and projected into a shared representation space, enabling the model to reason alongside natural language instructions. The model is particularly adept at tasks such as prediction, classification, site localization, quantitative analysis, and structure inference.
The architecture also incorporates modality-specific projectors that map diverse scientific data into the input space of a shared backbone. This allows Sage S2.1 Flash to process these inputs alongside text and images, making it highly versatile in multimodal reasoning and analysis.
Performance Across Benchmarks
The model was evaluated on 33 benchmarks across seven domains, achieving notable results:
- First place on OfficeQA Pro (72.2), MLE-bench Lite (90.3 HumanRank), Harvey LAB, and SpreadsheetBench~2.
- Second place on OSWorld-Verified (85.2), WebArena-Verified (67.6), and DeepResearch Bench~II (54.2).
These results highlight the model's strengths in using documentary evidence, producing verifiable workbooks, and executing machine learning tasks. It also demonstrated competitive performance in workflows requiring visual feedback, such as desktop and browser-based tasks.
Interactive and Verifiable Scientific Workflows
Sage S2.1 Flash is designed to operate in interactive environments, enabling it to perform tool use, search, and coding tasks. The model can utilize specialized software to conduct analyses, verify results against intermediate outputs, and adjust its approach based on visual feedback. This makes it particularly suited for complex scientific workflows that demand iterative analysis and verification.
Open-Source Availability
The model and its implementation are available on GitHub, allowing researchers and developers to explore its capabilities. Sage S2.1 Flash supports deployment through platforms like Hugging Face and ModelScope, with options for FP8 precision models.
With its advanced multimodal capabilities and strong benchmark performance, Sage S2.1 Flash represents a significant step forward in AI-driven scientific research and analysis.