Step-3.5-Flash model card and config (huggingface.co/stepfun-ai)

Step 3.5 Flash states 196.81B total (196B backbone + 0.81B head) with 11B activated, Top-8 of 288 experts, and 3 MTP layers. config.json shows the interleaved sliding-window pattern directly: layer_types has 12 full_attention + 36 sliding_attention, sliding_window 512, num_attention_groups 8, att_impl_type GQA, and per-layer rope_theta (5000000 on full-attention layers, 10000 elsewhere).