MiniMax-M2 and M2.7 model cards and configs (huggingface.co/MiniMaxAI)

MiniMax-M2 card states 230B total / 10B active and argues explicitly that keeping activations near 10B is the point ('10B activations = responsive agent loops + better unit economics'). Note this is a deliberate reversal of MiniMax-M1's 45.9B active: M2 is all-full attention (config attn_type_list is 1 for all 62 layers) with 256 local experts and 8 per token, whereas M1 used hybrid lightning attention. M2 and M2.7 ship under a NON-COMMERCIAL LICENSE (MIT-style terms, commercial use requires written authorization), which differs from M1's Apache-2.0. M2.7 (2026-04-09) raises max_position_embeddings to 204800 from M2's 196608 and keeps the identical 62/256/8 architecture.