gemma-4-26B-A4B-it config.json (text_config)

gemma4. text_config: num_hidden_layers 30, hidden_size 2816, num_attention_heads 16, num_key_value_heads 8, head_dim 256, global_head_dim 512, num_global_key_value_heads 2, sliding_window 1024, attention_k_eq_v true. layer_types over 30 layers = 25 sliding_attention + 5 full_attention (0-based 5,11,17,23,29). enable_moe_block TRUE, num_experts 128, top_k_experts 8, moe_intermediate_size 704, final_logit_softcapping 30.0. NO n_shared_experts field at all -> no always-on shared expert. rope_parameters full_attention proportional theta 1e6 partial_rotary 0.25 / sliding_attention default theta 1e4. max_position_embeddings 262144. Read raw 2026-10-04.