deepseek-ai/DeepSeek-V4-Pro-0813 config.json

Fetched 2026-10-18. model_type deepseek_v4, architectures DeepseekV4ForCausalLM, quantization_config quant_method fp8. Identical geometry to the preview config recorded under sources/mbl-cfg-deepseek-v4-pro: 61 layers, hidden_size 7168, 128 attention heads, num_key_value_heads 1, head_dim 512, qk_rope_head_dim 64, q_lora_rank 1536, no kv_lora_rank key. MoE: n_routed_experts 384, n_shared_experts 1, num_experts_per_tok 6, moe_intermediate_size 3072, routed_scaling_factor 2.5, topk_method noaux_tc, scoring_func sqrtsoftplus. compress_ratios length 64 (61 layers plus 3 MTP): 1 sliding (ratio 0), 30 CSA (ratio 4), 31 HCA (ratio 128) over the 61 layers, sliding_window 128, index_n_heads 64, index_head_dim 128, index_topk 1024. max_position_embeddings 1048576 with rope_scaling YaRN factor 16 over original_max_position_embeddings 65536. vocab_size 129280, num_nextn_predict_layers 1. This file is why the record copies the preview's KV verdict rather than deriving a new number.