StepFun Step 5 Preview model page

Fetched 2026-10-18. Vendor announcement and benchmark page. States a sparse MoE with 600B total parameters and 27B active per token, 1M-token context, vision input, and says the weights would be released on October 15, 2026. The comparison table is a vendor claim run at the High reasoning setting. Rows used here: GPQA Diamond 93.5% (GLM-5.3 Max 91.7%, Kimi K3 Max 93.5%), Terminal-Bench v2.1 85.0% (GLM-5.3 Max 83.9%). The page also lists DeepSWE v1.1 67.7 and many internal or non-standard benches. The Hugging Face card points at this page for methodology and says the GPQA and Terminal-Bench numbers above are the High setting. Not independently re-run.