MiniMax · 2026-06 · Sparse MoE · Multimodal
MiniMax-M3
MiniMax-M3 is a grouped-query attention (GQA) transformer released by MiniMax in 2026-06, with 60 layers, hidden size 6144 and a context window of 1,048,576 tokens.
Layer stack
Attention ×60 · Layers 60
Key facts
| Family | MiniMax |
|---|---|
| Released | 2026-06 |
| Params | 443.4 B |
| Active | 25.5 B |
| Context | 1,048,576 tokens |
| Attention | GQA (64:4) |
| Layers | 60 |
| Hidden | 6,144 |
| Heads | 64 |
| Vocab | 200,064 |
| Position | RoPE |
| Norm | RMSNorm |
| Activation | swigluoai |
| Dtype | bf16 |
| Architecture class | MiniMaxM3SparseForConditionalGeneration |
Architecture overview
Attention
grouped-query attention (GQA) — 64 q-heads / 4 kv-heads. RoPE θ=5,000,000.
Feed-forward / MoE
It is a sparse mixture-of-experts with 128 routed experts and top-4 routing; about 25.5B parameters are active per token.
The configuration ships 1 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-M2.7
| Model | MiniMax-M3 | MiniMax-M2.7 | Ratio |
|---|---|---|---|
| model_type | minimax_m3_vl | minimax_m2 | ≠ |
| architectures | MiniMaxM3SparseForConditionalGeneration | MiniMaxM2ForCausalLM | ≠ |
| hidden_size | 6144 | 3072 | ×2.00 |
| num_hidden_layers | 60 | 62 | ×0.968 |
| num_attention_heads | 64 | 48 | ×1.33 |
| num_key_value_heads | 4 | 8 | ×0.500 |
| head_dim | 128 | 128 | ≈1 |
| intermediate_size | 3072 | 1536 | ×2.00 |
| hidden_act | swigluoai | silu | ≠ |
| num_experts_per_tok | 4 | 8 | ×0.500 |
| n_shared_experts | 1 | — | |
| rope_theta | 5000000 | 5000000 | ≈1 |
| max_position_embeddings | 1048576 | 204800 | ×5.12 |
| rms_norm_eps | 0.000001 | 0.000001 | ≈1 |
| vocab_size | 200064 | 200064 | ≈1 |
| torch_dtype | bfloat16 | — | |
| num_nextn_predict_layers | 1 | — | |
| image_grid_pinpoints | [(336, 336), (336, 672), (336, 1008), (336, 1344), (336, 1680), (336, 2016), (672, 336), (672, 672), (672, 1008), (672, 1344), (672, 1680), (672, 2016), (1008, 336), (1008, 672), (1008, 1008), (1008, 1344), (1008, 1680), (1008, 2016), (1344, 336), (1344, 672), (1344, 1008), (1344, 1344), (1344, 1680), (1344, 2016), (1680, 336), (1680, 672), (1680, 1008), (1680, 1344), (1680, 1680), (1680, 2016), (2016, 336), (2016, 672), (2016, 1008), (2016, 1344), (2016, 1680), (2016, 2016)] | — | |
| image_seq_length | 576 | — | |
| image_token_index | 200025 | — | |
| video_token_index | 200026 | — | |
| multimodal_projector_bias | true | — | |
| num_reward_heads | 0 | — | |
| process_image_mode | dynamic_res | — | |
| projector_hidden_act | gelu | — | |
| vision_feature_layer | -1 | — |
Most similar architectures
- MiniMax-M3 MXFP8 (MiniMax, 2026-06) 1.000
- Qwen3.8 2.4T A95B (Qwen, 2026-08) 0.927
- MiniMax-Text-01 (MiniMax, 2025-01) 0.926
- MiniMax-M1 (MiniMax, 2025-06) 0.926
- GLM-4.6 (GLM, 2025-10) 0.924
- GLM-4.5 (GLM, 2025-07) 0.922
Similarity values range from 0 (no shared categorical features) to 1 (identical profiles).
Raw config fields
46 fields
| architectures | MiniMaxM3SparseForConditionalGeneration |
|---|---|
| model_type | minimax_m3_vl |
| image_grid_pinpoints | [(336, 336), (336, 672), (336, 1008), (336, 1344), (336, 1680), (336, 2016), (672, 336), (672, 672), (672, 1008), (672, 1344), (672, 1680), (672, 2016), (1008, 336), (1008, 672), (1008, 1008), (1008, 1344), (1008, 1680), (1008, 2016), (1344, 336), (1344, 672), (1344, 1008), (1344, 1344), (1344, 1680), (1344, 2016), (1680, 336), (1680, 672), (1680, 1008), (1680, 1344), (1680, 1680), (1680, 2016), (2016, 336), (2016, 672), (2016, 1008), (2016, 1344), (2016, 1680), (2016, 2016)] |
| image_seq_length | 576 |
| image_token_index | 200025 |
| video_token_index | 200026 |
| multimodal_projector_bias | true |
| num_reward_heads | 0 |
| process_image_mode | dynamic_res |
| projector_hidden_act | gelu |
| vision_feature_layer | -1 |
| vision_feature_select_strategy | full |
| torch_dtype | bfloat16 |
| transformers_version | 4.52.4 |
| projector_hidden_size | 6144 |
| hidden_size | 6144 |
| intermediate_size | 3072 |
| num_hidden_layers | 60 |
| num_attention_heads | 64 |
| num_key_value_heads | 4 |
| head_dim | 128 |
| vocab_size | 200064 |
| max_position_embeddings | 1048576 |
| rms_norm_eps | 0.000001 |
| use_gemma_norm | true |
| attention_output_gate | false |
| rope_theta | 5000000 |
| rotary_dim | 64 |
| partial_rotary_factor | 0.5 |
| hidden_act | swigluoai |
| use_qk_norm | true |
| tie_word_embeddings | false |
| dense_intermediate_size | 12288 |
| shared_intermediate_size | 3072 |
| num_local_experts | 128 |
| num_experts_per_tok | 4 |
| n_shared_experts | 1 |
| scoring_func | sigmoid |
| use_routing_bias | true |
| moe_layer_freq | 0×3 + 1×57 |
| qk_norm_type | per_head |
| num_mtp_modules | 7 |
| num_nextn_predict_layers | 1 |
| swiglu_alpha | 1.702 |
| swiglu_limit | 7 |
| routed_scaling_factor | 2 |