Kimi · 2026-07 · Sparsames MoE · Multimodal
Kimi K3
Kimi K3 ist ein Multi-Head-Latent-Attention (MLA)-Transformer, veröffentlicht von Kimi im 2026-07, mit 93 Schichten, Hidden-Größe 7168 und einem Kontextfenster von 1,048,576 Tokens.
Schichtstapel
Attention ×93 · Schichten 93
Kernfakten
| Labor | Kimi |
|---|---|
| Veröffentlicht | 2026-07 |
| Parameter | 2,780 Mrd. |
| Aktiv | 103 Mrd. |
| Kontext | 1,048,576 tokens |
| Attention | MLA · kv_lora_rank 512 |
| Schichten | 93 |
| Hidden-Dim. | 7,168 |
| Attention-Köpfe | 96 |
| Vokabular | 163,840 |
| Positionscodierung | learned/absolute |
| Normierung | RMSNorm |
| Aktivierung | situ |
| Präzision | 专家 MXFP4(attn/shared/mlp/lm_head 高精度) |
| Architekturklasse | KimiK3ForConditionalGeneration |
Architekturüberblick
Attention
Multi-Head-Latent-Attention (MLA). kv_lora_rank=512.
FFN / MoE
Es handelt sich um ein sparsames Mixture-of-Experts mit 896 gerouteten Experten und Top-16-Routing; etwa 103 Mrd. Parameter sind pro Token aktiv.
Feldweiser Vergleich
Verglichen mit dem vorherigen Modell desselben Labors; fehlt dieses, mit dem strukturell nächsten Modell. Die Spalte Ratio = dieses Modell ÷ Vergleichsmodell.
Vollfeldvergleich mit dem Vorgänger Kimi K2.6
| Modell | Kimi K3 | Kimi K2.6 | Ratio |
|---|---|---|---|
| model_type | kimi_k3 | kimi_k25 | ≠ |
| architectures | KimiK3ForConditionalGeneration | KimiK25ForConditionalGeneration | ≠ |
| hidden_size | 7168 | 7168 | ≈1 |
| num_hidden_layers | 93 | 61 | ×1.52 |
| num_attention_heads | 96 | 64 | ×1.50 |
| num_key_value_heads | 96 | 64 | ×1.50 |
| intermediate_size | 33792 | 18432 | ×1.83 |
| hidden_act | situ | silu | ≠ |
| n_routed_experts | — | 384 | |
| num_experts_per_tok | — | 8 | |
| n_shared_experts | — | 1 | |
| moe_intermediate_size | 3072 | 2048 | ×1.50 |
| kv_lora_rank | 512 | 512 | ≈1 |
| qk_rope_head_dim | 64 | 64 | ≈1 |
| q_lora_rank | 1536 | 1536 | ≈1 |
| rope_theta | — | 50000 | |
| max_position_embeddings | 1048576 | 262144 | ×4.00 |
| rms_norm_eps | 0.00001 | 0.00001 | ≈1 |
| vocab_size | 163840 | 163840 | ≈1 |
| num_nextn_predict_layers | 0 | 0 | ＝ |
| bos_token_id | 163584 | 163584 | ≈1 |
| dtype | bfloat16 | bfloat16 | ＝ |
| eos_token_id | 163586 | 163586 | ≈1 |
| ignore_index | -100 | -100 | ≈1 |
| image_placeholder | <|kimi_image_placeholder|> | — | |
| media_placeholder_token_id | 163605 | 163605 | ≈1 |
Strukturell ähnlichste Modelle
- Hunyuan 4 Preview (Hunyuan, 2026-08) 0.858
- GLM-5.2 (GLM, 2026-06) 0.858
- GLM-5.3 (GLM, 2026-08) 0.858
- GLM-5 (GLM, 2026-02) 0.853
- GLM-5.1 (GLM, 2026-04) 0.853
- GLM-5.3-Flash (GLM, 2026-08) 0.836
Ähnlichkeitswerte reichen von 0 (keine gemeinsamen kategorialen Merkmale) bis 1 (identische Profile).
Rohe Config-Felder
79 Felder
| architectures | KimiK3ForConditionalGeneration |
|---|---|
| bos_token_id | 163584 |
| dtype | bfloat16 |
| eos_token_id | 163586 |
| ignore_index | -100 |
| image_placeholder | <|kimi_image_placeholder|> |
| media_placeholder_token_id | 163605 |
| model_type | kimi_k3 |
| pad_token_id | 163839 |
| tie_word_embeddings | false |
| _name_or_path | |
| activation_situ_beta | 4 |
| activation_situ_linear_beta | 25 |
| add_cross_attention | false |
| attn_res_block_size | 12 |
| chunk_size_feed_forward | 0 |
| diversity_penalty | 0 |
| do_sample | false |
| early_stopping | false |
| encoder_no_repeat_ngram_size | 0 |
| first_k_dense_replace | 1 |
| hidden_act | situ |
| hidden_size | 7168 |
| initializer_range | 0.02 |
| intermediate_size | 33792 |
| is_decoder | false |
| is_encoder_decoder | false |
| kv_lora_rank | 512 |
| latent_moe_use_norm | true |
| length_penalty | 1 |
| max_length | 20 |
| max_position_embeddings | 1048576 |
| min_length | 0 |
| mla_use_nope | true |
| mla_use_output_gate | true |
| moe_intermediate_size | 3072 |
| moe_layer_freq | 1 |
| moe_renormalize | true |
| moe_router_activation_func | sigmoid |
| no_repeat_ngram_size | 0 |
| num_attention_heads | 96 |
| num_beam_groups | 1 |
| num_beams | 1 |
| num_expert_group | 1 |
| num_experts | 896 |
| num_experts_per_token | 16 |
| num_hidden_layers | 93 |
| num_key_value_heads | 96 |
| num_nextn_predict_layers | 0 |
| num_return_sequences | 1 |
| num_shared_experts | 2 |
| output_attentions | false |
| output_hidden_states | false |
| output_scores | false |
| q_lora_rank | 1536 |
| qk_nope_head_dim | 128 |
| qk_rope_head_dim | 64 |
| remove_invalid_values | false |
| repetition_penalty | 1 |
| return_dict | true |
| return_dict_in_generate | false |
| rms_norm_eps | 0.00001 |
| routed_expert_hidden_size | 3584 |
| routed_scaling_factor | 1 |
| temperature | 1 |
| tf_legacy_loss | false |
| tie_encoder_decoder | false |
| top_k | 50 |
| top_p | 1 |
| topk_group | 1 |
| topk_method | noaux_tc |
| torchscript | false |
| transformers_version | 4.56.2 |
| typical_p | 1 |
| use_bfloat16 | false |
| use_cache | true |
| use_grouped_topk | true |
| v_head_dim | 128 |
| vocab_size | 163840 |