Kimi · 2025-09 · Sparsames MoE
Kimi K2 0905
Kimi K2 0905 ist ein Multi-Head-Latent-Attention (MLA)-Transformer, veröffentlicht von Kimi im 2025-09, mit 61 Schichten, Hidden-Größe 7168 und einem Kontextfenster von 262,144 Tokens.
Schichtstapel
Attention ×61 · Schichten 61
Kernfakten
| Labor | Kimi |
|---|---|
| Veröffentlicht | 2025-09 |
| Parameter | 1,040.2 Mrd. |
| Aktiv | 32.8 Mrd. |
| Kontext | 262,144 tokens |
| Attention | MLA · kv_lora_rank 512 |
| Schichten | 61 |
| Hidden-Dim. | 7,168 |
| Attention-Köpfe | 64 |
| Vokabular | 163,840 |
| Positionscodierung | RoPE |
| Normierung | RMSNorm |
| Aktivierung | silu |
| Präzision | fp8 |
| Architekturklasse | DeepseekV3ForCausalLM |
Architekturüberblick
Attention
Multi-Head-Latent-Attention (MLA). kv_lora_rank=512. RoPE θ=50,000.
FFN / MoE
Es handelt sich um ein sparsames Mixture-of-Experts mit 384 gerouteten Experten und Top-8-Routing; etwa 32.8 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 Base
| Modell | Kimi K2 0905 | Kimi K2 Base | Ratio |
|---|---|---|---|
| model_type | kimi_k2 | kimi_k2 | ＝ |
| architectures | DeepseekV3ForCausalLM | DeepseekV3ForCausalLM | ＝ |
| hidden_size | 7168 | 7168 | ≈1 |
| num_hidden_layers | 61 | 61 | ≈1 |
| num_attention_heads | 64 | 64 | ≈1 |
| num_key_value_heads | 64 | 64 | ≈1 |
| intermediate_size | 18432 | 18432 | ≈1 |
| hidden_act | silu | silu | ＝ |
| n_routed_experts | 384 | 384 | ≈1 |
| num_experts_per_tok | 8 | 8 | ≈1 |
| n_shared_experts | 1 | 1 | ≈1 |
| moe_intermediate_size | 2048 | 2048 | ≈1 |
| kv_lora_rank | 512 | 512 | ≈1 |
| qk_rope_head_dim | 64 | 64 | ≈1 |
| q_lora_rank | 1536 | 1536 | ≈1 |
| rope_theta | 50000 | 50000 | ≈1 |
| max_position_embeddings | 262144 | 131072 | ×2.00 |
| rms_norm_eps | 0.00001 | 0.000001 | ×10.00 |
| vocab_size | 163840 | 163840 | ≈1 |
| torch_dtype | bfloat16 | bfloat16 | ＝ |
| num_nextn_predict_layers | 0 | 0 | ＝ |
| _attn_implementation_autoset | false | — | |
| _name_or_path | | — | |
| add_cross_attention | false | — | |
| attention_bias | false | false | ＝ |
| attention_dropout | 0 | 0 | ＝ |
Strukturell ähnlichste Modelle
- Kimi K2.5 (Kimi, 2026-02) 1.000
- Kimi K2.6 (Kimi, 2026-04) 1.000
- Kimi K2 Base (Kimi, 2025-07) 1.000
- DeepSeek-V3 (DeepSeek, 2024-12) 1.000
- DeepSeek-V3 Base (DeepSeek, 2024-12) 1.000
- DeepSeek-R1 (DeepSeek, 2025-01) 1.000
Ähnlichkeitswerte reichen von 0 (keine gemeinsamen kategorialen Merkmale) bis 1 (identische Profile).
Rohe Config-Felder
76 Felder
| _attn_implementation_autoset | false |
|---|---|
| _name_or_path | |
| add_cross_attention | false |
| architectures | DeepseekV3ForCausalLM |
| attention_bias | false |
| attention_dropout | 0 |
| aux_loss_alpha | 0.001 |
| bos_token_id | 163584 |
| chunk_size_feed_forward | 0 |
| diversity_penalty | 0 |
| do_sample | false |
| early_stopping | false |
| encoder_no_repeat_ngram_size | 0 |
| eos_token_id | 163585 |
| ep_size | 1 |
| first_k_dense_replace | 1 |
| hidden_act | silu |
| hidden_size | 7168 |
| initializer_range | 0.02 |
| intermediate_size | 18432 |
| is_decoder | false |
| is_encoder_decoder | false |
| kv_lora_rank | 512 |
| length_penalty | 1 |
| max_length | 20 |
| max_position_embeddings | 262144 |
| min_length | 0 |
| model_type | kimi_k2 |
| moe_intermediate_size | 2048 |
| moe_layer_freq | 1 |
| n_group | 1 |
| n_routed_experts | 384 |
| n_shared_experts | 1 |
| no_repeat_ngram_size | 0 |
| norm_topk_prob | true |
| num_attention_heads | 64 |
| num_beam_groups | 1 |
| num_beams | 1 |
| num_experts_per_tok | 8 |
| num_hidden_layers | 61 |
| num_key_value_heads | 64 |
| num_nextn_predict_layers | 0 |
| num_return_sequences | 1 |
| output_attentions | false |
| output_hidden_states | false |
| output_scores | false |
| pad_token_id | 163839 |
| pretraining_tp | 1 |
| 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 |
| rope_theta | 50000 |
| routed_scaling_factor | 2.827 |
| scoring_func | sigmoid |
| seq_aux | true |
| temperature | 1 |
| tf_legacy_loss | false |
| tie_encoder_decoder | false |
| tie_word_embeddings | false |
| top_k | 50 |
| top_p | 1 |
| topk_group | 1 |
| topk_method | noaux_tc |
| torch_dtype | bfloat16 |
| torchscript | false |
| transformers_version | 4.51.3 |
| typical_p | 1 |
| use_bfloat16 | false |
| use_cache | true |
| v_head_dim | 128 |
| vocab_size | 163840 |