MiniMax · 2026-06 · 稀疏 MoE · 多模态
MiniMax-M3 MXFP8
MiniMax-M3 MXFP8 是 MiniMax 于 2026-06 发布的 分组查询注意力（GQA） Transformer，共 60 层，隐藏维度 6144，上下文窗口 1,048,576 个 token。
层堆栈
注意力 ×60 · 层数 60
关键参数
| 实验室 | MiniMax |
|---|---|
| 发布时间 | 2026-06 |
| 总参数 | 443.4 B |
| 激活参数 | 25.5 B |
| 上下文 | 1,048,576 tokens |
| 注意力 | GQA (64:4) |
| 层数 | 60 |
| 隐藏维度 | 6,144 |
| 注意力头 | 64 |
| 词表 | 200,064 |
| 位置编码 | RoPE |
| 归一化 | RMSNorm |
| 激活函数 | swigluoai |
| 训练精度 | MXFP8 |
| 架构类名 | MiniMaxM3SparseForConditionalGeneration |
架构概览
注意力
分组查询注意力（GQA） — 64 q-heads / 4 kv-heads. RoPE θ=5,000,000.
前馈 / MoE
它是稀疏混合专家模型：128 个路由专家、top-4 路由；每个 token 约激活 25.5B 参数。
配置中包含 1 个额外的多 token 预测（MTP）层，用于加速解码。
字段级对比
对比对象为同实验室的上一代模型；无前代时取结构最相似的模型。倍数列 = 本模型数值 ÷ 对比模型数值。
与前代 MiniMax-M3 的全字段对比
| 模型 | MiniMax-M3 MXFP8 | MiniMax-M3 | 倍数 |
|---|---|---|---|
| model_type | minimax_m3_vl | minimax_m3_vl | ＝ |
| architectures | MiniMaxM3SparseForConditionalGeneration | MiniMaxM3SparseForConditionalGeneration | ＝ |
| hidden_size | 6144 | 6144 | ≈1 |
| num_hidden_layers | 60 | 60 | ≈1 |
| num_attention_heads | 64 | 64 | ≈1 |
| num_key_value_heads | 4 | 4 | ≈1 |
| head_dim | 128 | 128 | ≈1 |
| intermediate_size | 3072 | 3072 | ≈1 |
| hidden_act | swigluoai | swigluoai | ＝ |
| num_experts_per_tok | 4 | 4 | ≈1 |
| n_shared_experts | 1 | 1 | ≈1 |
| rope_theta | 5000000 | 5000000 | ≈1 |
| max_position_embeddings | 1048576 | 1048576 | ≈1 |
| rms_norm_eps | 0.000001 | 0.000001 | ≈1 |
| vocab_size | 200064 | 200064 | ≈1 |
| torch_dtype | bfloat16 | 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)] | [(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 | 576 | ≈1 |
| image_token_index | 200025 | 200025 | ≈1 |
| video_token_index | 200026 | 200026 | ≈1 |
| multimodal_projector_bias | true | true | ＝ |
| num_reward_heads | 0 | 0 | ＝ |
| process_image_mode | dynamic_res | dynamic_res | ＝ |
| projector_hidden_act | gelu | gelu | ＝ |
| vision_feature_layer | -1 | -1 | ≈1 |
结构最相似的模型
- MiniMax-M3 (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
相似度取值范围 0（无共同类别特征）到 1（结构画像完全一致）。
原始 config 字段
46 个字段
| 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 |
| dtype | bfloat16 |
| 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 | 1 |
| swiglu_alpha | 1.702 |
| swiglu_limit | 7 |
| routed_scaling_factor | 2 |