Qwen3.5-27B config.json (Hugging Face, Qwen/Qwen3.5-27B)

Raw config.json read 2026-10-18 from the Qwen/Qwen3.5-27B repo. text_config: num_hidden_layers 64, hidden_size 5120, num_attention_heads 24, num_key_value_heads 4, head_dim 256, full_attention_interval 4, dtype bfloat16. layer_types is 48 linear_attention + 16 full_attention, full layers at 0-based indices 3,7,11,15,19,23,27,31,35,39,43,47,51,55,59,63. Linear layers declare linear_num_key_heads 16, linear_num_value_heads 48, linear_key_head_dim 128, linear_value_head_dim 128, linear_conv_kernel_dim 4 — a fixed recurrent state, not a cache that grows with context. max_position_embeddings 262144, rope_parameters.rope_type default with no scaling factor. Dense: no num_experts key. The catalog has no model record for this checkpoint; this source is the geometry the KV record derives from.