MiniMaxAI/MiniMax-M2 config.json

minimax_m2. num_hidden_layers 62, hidden_size 3072, num_attention_heads 48, num_key_value_heads 8, head_dim 128, rope_theta 5000000, rope_scaling null, sliding_window null, max_position_embeddings 196608, rotary_dim 64. MoE: num_local_experts 256, num_experts_per_tok 8, shared_intermediate_size 0 (no always-on shared expert). Read raw 2026-10-04 - NOTE the record covers the M2 line whose context lengths differ per release: M2 196608, M2.1 196608, M2.7 204800, so the record's 204800 is the M2.7 figure and this config documents M2.