Sage S2.1 Flash Technical Report
The authors report four first-place finishes against five baselines for a 320-billion-parameter mixture-of-experts model.
The abstract calls the model Sage S1.5 Flash, despite the report’s S2.1 title. It combines text and images with dedicated representations for molecular graphs, biological sequences and spectra. Agent training uses verifiable environments, with completion checked through execution results, environment changes or submitted artifacts. Across 33 benchmarks in seven domains, the reported wins include OfficeQA Pro and MLE-bench Lite; it places second on OSWorld-Verified and WebArena-Verified.
HF Daily Papers' note
The abstract calls the model Sage S1.5 Flash, despite the report’s S2.1 title. It combines text and images with dedicated representations for molecular graphs, biological sequences and spectra. Agent training uses verifiable environments, with completion checked through execution results, environment changes or submitted artifacts. Across 33 benchmarks in seven domains, the reported wins include OfficeQA Pro and MLE-bench Lite; it places second on OSWorld-Verified and WebArena-Verified.
HF Daily Papers' note
score 5