Ling-3.0-flash model card and config (huggingface.co/inclusionAI)

Ling-3.0-flash card: 124B total / 5.1B active, native hybrid linear attention from pretraining onward with '5:1 alternating stacking of Kimi Delta Attention (KDA) and MLA', 512 routed + 1 shared expert, 8 activated, expert intermediate size 768. MIT licensed. config.json (model_type bailing_hybrid) has kv_lora_rank 512, qk_nope_head_dim 128, qk_rope_head_dim 64, v_head_dim 128, max_window_layers 20, layer_group_size 6, num_shared_experts 1, max_position_embeddings 262144. The card's 124B differs from the 127.49B safetensors total.