Step3 — Retired model
Step3 is a cutting-edge multimodal reasoning model—built on a Mixture-of-Experts architecture with 321B total parameters and 38B active. It is designed end-to-end to minimize decoding costs while delivering top-tier performance in vision–language reasoning. Through the co-design of Multi-Matrix Factorization Attention (MFA) and Attention-FFN Disaggregation (AFD), Step3 maintains exceptional efficiency across both flagship and low-end accelerators.
Technical specifications
Frequently Asked Questions
Can I still use the retired Step3 model?
No. Step3 has been retired and is retained here only as a historical reference. See the active alternatives linked on this page.
Relay integrity checks
No integrity-check data is available for this model yet.
BazaarLink Probe verifies endpoints that claim to serve this model and flags behavior that differs from the reference baseline.
View the full integrity report →Related models and links
Related models and links