StepFun releases Step 3.7 Flash for efficient inference

阶跃星辰Step 3.7 Flash发布，专为高效推理设计

StepFun released Step 3.7 Flash with a 196B MoE architecture, using multi-matrix factorized attention to cut KV-cache cost to about 22% of DeepSeek models.

Why it matters: HKR-H/K/R all pass: Step 3.7 Flash has concrete specs, not just launch copy, with 196B MoE and ~22% KV-cache cost versus DeepSeek. It is below top-lab flagship weight, so 78 featured.