mistralai/Mamba-Codestral-7B-v0.1 config.json

mamba2, Mamba2ForCausalLM. num_hidden_layers 64, hidden_size 4096, head_dim 64 (the Mamba d_state/inner width convention, NOT attention head_dim). NO num_attention_heads, NO num_key_value_heads, no max_position_embeddings -> the model has no attention layers, so KV cache, GQA ratio and context_scaling are all undefined rather than zero. Read raw 2026-10-04.