Qwen3.5-9B config.json (text_config, official repo)

Read raw from the official Qwen/Qwen3.5-9B repo on 2026-10-18. text_config: num_hidden_layers 32, full_attention_interval 4, layer_types is 24 linear_attention plus 8 full_attention (every 4th layer), num_attention_heads 16, num_key_value_heads 4, head_dim 256, max_position_embeddings 262144, dtype bfloat16. Linear-attention state: linear_num_key_heads 16, linear_num_value_heads 32, linear_key_head_dim 128, linear_value_head_dim 128, linear_conv_kernel_dim 4. The file is the vendors own config, not a mirror.