MiniMax · 2025-10 · Sparse MoE
MiniMax-M2
MiniMax-M2 is a grouped-query attention (GQA) transformer released by MiniMax in 2025-10, with 62 layers, hidden size 3072 and a context window of 196,608 tokens.
Layer stack
Attention ×62 · Layers 62
Key facts
| Family | MiniMax |
|---|---|
| Released | 2025-10 |
| Params | 228.3 B |
| Active | 10.6 B |
| Context | 196,608 tokens |
| Attention | GQA (48:8) |
| Layers | 62 |
| Hidden | 3,072 |
| Heads | 48 |
| Vocab | 200,064 |
| Position | RoPE |
| Norm | RMSNorm |
| Activation | silu |
| Dtype | fp8 |
| Architecture class | MiniMaxM2ForCausalLM |
Architecture overview
Attention
grouped-query attention (GQA) — 48 q-heads / 8 kv-heads. RoPE θ=5,000,000.
Feed-forward / MoE
It is a sparse mixture-of-experts with 256 routed experts and top-8 routing; about 10.6B parameters are active per token.
The configuration ships 3 extra multi-token-prediction (MTP) layers used to accelerate decoding.
Field-level comparison
Compared against the previous model of the same lab; where none exists, against the structurally closest model. The ratio column is this model divided by the comparison model.
Full-field comparison vs predecessor MiniMax-M1
| Model | MiniMax-M2 | MiniMax-M1 | Ratio |
|---|---|---|---|
| model_type | minimax_m2 | minimax | ≠ |
| architectures | MiniMaxM2ForCausalLM | MiniMaxForCausalLM | ≠ |
| hidden_size | 3072 | 6144 | ×0.500 |
| num_hidden_layers | 62 | 80 | ×0.775 |
| num_attention_heads | 48 | 64 | ×0.750 |
| num_key_value_heads | 8 | 8 | ≈1 |
| head_dim | 128 | 128 | ≈1 |
| intermediate_size | 1536 | 9216 | ×0.167 |
| hidden_act | silu | silu | ＝ |
| num_experts_per_tok | 8 | 2 | ×4.00 |
| rope_theta | 5000000 | 10000000 | ×0.500 |
| max_position_embeddings | 196608 | 10240000 | ×0.019 |
| rms_norm_eps | 0.000001 | 0.00001 | ×0.100 |
| vocab_size | 200064 | 200064 | ≈1 |
| attention_dropout | 0 | 0 | ＝ |
| attn_type_list | 1×62 | — | |
| initializer_range | 0.02 | 0.02 | ≈1 |
| layernorm_full_attention_beta | 1 | — | |
| layernorm_linear_attention_beta | 1 | — | |
| layernorm_mlp_beta | 1 | — | |
| mlp_intermediate_size | 8192 | — | |
| mtp_transformer_layers | 1 | — | |
| num_local_experts | 256 | 32 | ×8.00 |
| num_mtp_modules | 3 | — | |
| output_router_logits | false | false | ＝ |
| qk_norm_type | per_layer | — |
Most similar architectures
- MiniMax-M2.5 (MiniMax, 2026-02) 1.000
- MiniMax-M2.7 (MiniMax, 2026-04) 1.000
- Qwen3.8-Flash-Next (Qwen, 2026-08) 0.934
- GLM-4.5-Air (GLM, 2025-07) 0.934
- Qwen3-Next 80B A3B (Qwen, 2025-09) 0.933
- Qwen3-Coder-Next (Qwen, 2026-01) 0.933
Similarity values range from 0 (no shared categorical features) to 1 (identical profiles).
Raw config fields
38 fields
| architectures | MiniMaxM2ForCausalLM |
|---|---|
| attention_dropout | 0 |
| attn_type_list | 1×62 |
| head_dim | 128 |
| hidden_act | silu |
| hidden_size | 3072 |
| initializer_range | 0.02 |
| intermediate_size | 1536 |
| layernorm_full_attention_beta | 1 |
| layernorm_linear_attention_beta | 1 |
| layernorm_mlp_beta | 1 |
| max_position_embeddings | 196608 |
| mlp_intermediate_size | 8192 |
| model_type | minimax_m2 |
| mtp_transformer_layers | 1 |
| num_attention_heads | 48 |
| num_experts_per_tok | 8 |
| num_hidden_layers | 62 |
| num_key_value_heads | 8 |
| num_local_experts | 256 |
| num_mtp_modules | 3 |
| output_router_logits | false |
| qk_norm_type | per_layer |
| rms_norm_eps | 0.000001 |
| rope_theta | 5000000 |
| rotary_dim | 64 |
| router_aux_loss_coef | 0.001 |
| router_jitter_noise | 0 |
| scoring_func | sigmoid |
| shared_intermediate_size | 0 |
| shared_moe_mode | sigmoid |
| tie_word_embeddings | false |
| transformers_version | 4.57.1 |
| use_cache | true |
| use_mtp | true |
| use_qk_norm | true |
| use_routing_bias | true |
| vocab_size | 200064 |