Kimi · 2025-10 · Sparse MoE
Kimi Linear 48B A3B
Kimi Linear 48B A3B is a multi-head latent attention (MLA) transformer released by Kimi in 2025-10, with 27 layers, hidden size 2304 and a context window of 262,144 tokens.
Layer stack
Attention ×27 · Layers 27
Key facts
| Family | Kimi |
|---|---|
| Released | 2025-10 |
| Params | 50 B |
| Active | 2.8 B |
| Context | 262,144 tokens |
| Attention | MLA · kv_lora_rank 512 |
| Layers | 27 |
| Hidden | 2,304 |
| Heads | 32 |
| Vocab | 163,840 |
| Position | RoPE |
| Norm | RMSNorm |
| Activation | silu |
| Dtype | bf16 |
| Architecture class | KimiLinearForCausalLM |
Architecture overview
Attention
multi-head latent attention (MLA). kv_lora_rank=512. RoPE θ=10,000.
Feed-forward / MoE
It is a sparse mixture-of-experts with 256 routed experts and top-8 routing; about 2.8B parameters are active per token.
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 Kimi K2 0905
| Model | Kimi Linear 48B A3B | Kimi K2 0905 | Ratio |
|---|---|---|---|
| model_type | kimi_linear | kimi_k2 | ≠ |
| architectures | KimiLinearForCausalLM | DeepseekV3ForCausalLM | ≠ |
| hidden_size | 2304 | 7168 | ×0.321 |
| num_hidden_layers | 27 | 61 | ×0.443 |
| num_attention_heads | 32 | 64 | ×0.500 |
| num_key_value_heads | 32 | 64 | ×0.500 |
| head_dim | 72 | — | |
| intermediate_size | 9216 | 18432 | ×0.500 |
| hidden_act | silu | silu | ＝ |
| n_routed_experts | — | 384 | |
| num_experts_per_tok | — | 8 | |
| n_shared_experts | — | 1 | |
| moe_intermediate_size | 1024 | 2048 | ×0.500 |
| kv_lora_rank | 512 | 512 | ≈1 |
| qk_rope_head_dim | 64 | 64 | ≈1 |
| q_lora_rank | — | 1536 | |
| rope_theta | 10000 | 50000 | ×0.200 |
| max_position_embeddings | — | 262144 | |
| rms_norm_eps | 0.00001 | 0.00001 | ≈1 |
| vocab_size | 163840 | 163840 | ≈1 |
| torch_dtype | — | bfloat16 | |
| num_nextn_predict_layers | 0 | 0 | ＝ |
| bos_token_id | 163584 | 163584 | ≈1 |
| dtype | bfloat16 | — | |
| eos_token_id | 163586 | 163585 | ≈1 |
| first_k_dense_replace | 1 | 1 | ≈1 |
Most similar architectures
- DeepSeek-V2 Lite (DeepSeek, 2024-05) 0.996
- GLM-4.7-Flash (GLM, 2026-01) 0.995
- Moonlight 16B (Kimi, 2025-02) 0.992
- DeepSeek-V2 (DeepSeek, 2024-05) 0.990
- Kimi K2 0905 (Kimi, 2025-09) 0.923
- Kimi K2.5 (Kimi, 2026-02) 0.923
Similarity values range from 0 (no shared categorical features) to 1 (identical profiles).
Raw config fields
39 fields
| architectures | KimiLinearForCausalLM |
|---|---|
| bos_token_id | 163584 |
| dtype | bfloat16 |
| eos_token_id | 163586 |
| first_k_dense_replace | 1 |
| head_dim | 72 |
| hidden_act | silu |
| hidden_size | 2304 |
| initializer_range | 0.02 |
| intermediate_size | 9216 |
| kv_lora_rank | 512 |
| mla_use_nope | true |
| model_max_length | 1048576 |
| model_type | kimi_linear |
| moe_intermediate_size | 1024 |
| moe_layer_freq | 1 |
| moe_renormalize | true |
| moe_router_activation_func | sigmoid |
| num_attention_heads | 32 |
| num_expert_group | 1 |
| num_experts | 256 |
| num_experts_per_token | 8 |
| num_hidden_layers | 27 |
| num_key_value_heads | 32 |
| num_nextn_predict_layers | 0 |
| num_shared_experts | 1 |
| pad_token_id | 163839 |
| qk_nope_head_dim | 128 |
| qk_rope_head_dim | 64 |
| rms_norm_eps | 0.00001 |
| rope_theta | 10000 |
| routed_scaling_factor | 2.446 |
| tie_word_embeddings | false |
| topk_group | 1 |
| transformers_version | 4.57.1 |
| use_cache | true |
| use_grouped_topk | true |
| v_head_dim | 128 |
| vocab_size | 163840 |