Sage S2.1 Flash
AvailableShanghai AI Laboratory's scientific multimodal agent model (书生 S2.1 Flash), released and open-sourced Nov 16 2026 under the MIT License on Hugging Face and ModelScope (BF16 and FP8). About 320B-parameter MoE (288 routed experts, top-8; HF reports 321.3B) with a configured 1,048,576-token position limit. Uses Sage SciPerception projectors to take more than ten scientific input types (molecular graphs, crystal structures, protein sequences and 3D structures, DNA/RNA, spectra, mass spec, XRD, ECG, EEG, light curves) alongside text and images, plus search, coding, tool use and visual operation of scientific software. Lab says inference cost is about one-tenth of its Intern agent model; ranks first on MLE-bench Lite, OfficeQA Pro, Harvey LAB and SpreadsheetBench 2 in its own comparison (vendor-reported). Active parameters not stated.
Specifications
- License
- Open source · MIT
- Weights
- Downloadable
- Architecture
- Mixture-of-Experts
- Parameters
- 320B
- Context window
- 1.0M tokens
- Max output
- —
- Knowledge cutoff
- —
- Price (in / out, $/M)
- —
- Modalities
- TextVisionCode
Benchmarks
No benchmark scores recorded yet. Spotted some? Submit a correction.
Vendor-reported figures are claims until independently verified. See methodology.