MiniMax-M2
MiniMax · MiniMax M2 · open weights
- GQA
- RoPE
- Partial RoPE
- Pre-norm
- QK-norm
- MoE
- MTP
Facts and where they come from
| Released | 2025-10 | config.jsonconfig.jsonHugging Face repository creation date (api.createdAt) |
|---|---|---|
| Licence | other | config.jsonconfig.jsonREADME metadata: license |
| Total parameters | 230B | labmodel cardREADME: 230 billion total parameters with 10 billion active |
| Active parameters | 10B | labmodel cardREADME: 10 billion active |
| Context length | 192K tokens | config.jsonconfig.jsonmax_position_embeddings |
| Norm placement | pre | codemodelling codetransformers 5.18.0 / repo modelling code for minimax_m2: input_layernorm before attention, post_attention_layernorm before the MLP |
| Norm type | RMSNorm | codemodelling codetransformers 5.18.0 / repo modelling code for minimax_m2: input_layernorm before attention, post_attention_layernorm before the MLP |
| QK-norm | yes | config.jsonconfig.jsonuse_qk_norm |
| Positional encoding | RoPE on 50% of each head | config.jsonconfig.jsonrotary_dim / head_dim |
| Parallel attention and MLP | no | codemodelling codetransformers 5.18.0 / repo modelling code for minimax_m2: input_layernorm before attention, post_attention_layernorm before the MLP |
Architecture, drawn from the data
GQA 48q/8kv. Each column is one layer: its token mixer above, its feed-forward block below. Paler columns reuse another layer’s keys and values.
Modelled costs
From the cost model, batch size 1. Totals the lab states are in the table above; differences come from rounding, from what a lab counts, or from parts the model does not describe (listed on the about page).
| Parameters (modelled) | 229B |
|---|---|
| Active per token (modelled) | 11B |
| Without embeddings and output head | 227B total, 9.8B active |
| Multi-token-prediction layers (extra) | 11.1B |
| Published weights (Hugging Face count) | 229B |
| KV cache per token, BF16 (layers that grow with context) | 248 KiB |
| KV cache + state at 192K tokens, BF16 | 46.5 GiB |
| Decode FLOPs per token at 4K context | 27.1 GFLOP |
| Prefill FLOPs for a 4K prompt | 93.1 TFLOP |
KV cache against context
Every architecture field
| Field | Value | Source |
|---|---|---|
| d_model | 3,072 | config.jsonconfig.jsonhidden_size |
| vocab | 200,064 | config.jsonconfig.jsonvocab_size |
| tied_embeddings | false | config.jsonconfig.jsontie_word_embeddings |
| mixers.full.type | attn | codemodelling codeattention block |
| mixers.full.heads | 48 | config.jsonconfig.jsonnum_attention_heads |
| mixers.full.kv_heads | 8 | config.jsonconfig.jsonnum_key_value_heads |
| mixers.full.head_dim | 128 | config.jsonconfig.jsonhead_dim |
| mixers.full.qk_norm | true | config.jsonconfig.jsonuse_qk_norm |
| ffns.moe.type | moe | codemodelling codeMoE block |
| ffns.moe.experts | 256 | config.jsonconfig.jsonnum_local_experts |
| ffns.moe.active | 8 | config.jsonconfig.jsonnum_experts_per_tok |
| ffns.moe.d_expert | 1,536 | config.jsonconfig.jsonintermediate_size |
| ffns.moe.gated | true | codemodelling codeexperts are gated MLPs |
| layout | 62× full/moe | config.jsonconfig.jsonattn_type_list |
| mtp_layers | 3 | config.jsonconfig.jsonnum_mtp_modules |
Sources
- config.json @ 757303d
- model card
- modelling code · transformers 5.18.0 modelling code, or the model repository's own modelling file at the pinned revision
Listed in the LLM Architecture Gallery checklist as “MiniMax M2 (230B)” (name only; see about).