DeepSeek · 2024-05 · Sparse MoE
DeepSeek-V2
DeepSeek-V2 is a multi-head latent attention (MLA) transformer released by DeepSeek in 2024-05, with 60 layers, hidden size 5120 and a context window of 163,840 tokens.
Layer stack
Attention ×60 · Layers 60
Key facts
| Family | DeepSeek |
|---|---|
| Released | 2024-05 |
| Params | 236 B |
| Active | 21 B |
| Context | 163,840 tokens |
| Attention | MLA · kv_lora_rank 512 |
| Layers | 60 |
| Hidden | 5,120 |
| Heads | 128 |
| Vocab | 102,400 |
| Position | RoPE |
| Norm | RMSNorm |
| Activation | silu |
| Dtype | bf16 |
| Architecture class | DeepseekV2ForCausalLM |
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 160 routed experts and top-6 routing; about 21B 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 DeepSeekMoE 16B
| Model | DeepSeek-V2 | DeepSeekMoE 16B | Ratio |
|---|---|---|---|
| model_type | deepseek_v2 | deepseek | ≠ |
| architectures | DeepseekV2ForCausalLM | DeepseekForCausalLM | ≠ |
| hidden_size | 5120 | 2048 | ×2.50 |
| num_hidden_layers | 60 | 28 | ×2.14 |
| num_attention_heads | 128 | 16 | ×8.00 |
| num_key_value_heads | 128 | 16 | ×8.00 |
| intermediate_size | 12288 | 10944 | ×1.12 |
| hidden_act | silu | silu | ＝ |
| n_routed_experts | 160 | 64 | ×2.50 |
| num_experts_per_tok | 6 | 6 | ≈1 |
| n_shared_experts | 2 | 2 | ≈1 |
| moe_intermediate_size | 1536 | 1408 | ×1.09 |
| kv_lora_rank | 512 | — | |
| qk_rope_head_dim | 64 | — | |
| q_lora_rank | 1536 | — | |
| rope_theta | 10000 | 10000 | ≈1 |
| max_position_embeddings | 163840 | 4096 | ×40.00 |
| rms_norm_eps | 0.000001 | 0.000001 | ≈1 |
| vocab_size | 102400 | 102400 | ≈1 |
| torch_dtype | bfloat16 | bfloat16 | ＝ |
| attention_bias | false | false | ＝ |
| attention_dropout | 0 | 0 | ＝ |
| aux_loss_alpha | 0.001 | — | |
| bos_token_id | 100000 | 100000 | ≈1 |
| eos_token_id | 100001 | 100001 | ≈1 |
| first_k_dense_replace | 1 | 1 | ≈1 |
Most similar architectures
- GLM-4.7-Flash (GLM, 2026-01) 0.992
- Kimi Linear 48B A3B (Kimi, 2025-10) 0.990
- DeepSeek-V2 Lite (DeepSeek, 2024-05) 0.985
- Moonlight 16B (Kimi, 2025-02) 0.983
- Kimi K2 0905 (Kimi, 2025-09) 0.937
- Kimi K2.5 (Kimi, 2026-02) 0.937
Similarity values range from 0 (no shared categorical features) to 1 (identical profiles).
Raw config fields
41 fields
| architectures | DeepseekV2ForCausalLM |
|---|---|
| attention_bias | false |
| attention_dropout | 0 |
| aux_loss_alpha | 0.001 |
| bos_token_id | 100000 |
| eos_token_id | 100001 |
| first_k_dense_replace | 1 |
| hidden_act | silu |
| hidden_size | 5120 |
| initializer_range | 0.02 |
| intermediate_size | 12288 |
| kv_lora_rank | 512 |
| max_position_embeddings | 163840 |
| model_type | deepseek_v2 |
| moe_intermediate_size | 1536 |
| moe_layer_freq | 1 |
| n_group | 8 |
| n_routed_experts | 160 |
| n_shared_experts | 2 |
| norm_topk_prob | false |
| num_attention_heads | 128 |
| num_experts_per_tok | 6 |
| num_hidden_layers | 60 |
| num_key_value_heads | 128 |
| pretraining_tp | 1 |
| q_lora_rank | 1536 |
| qk_nope_head_dim | 128 |
| qk_rope_head_dim | 64 |
| rms_norm_eps | 0.000001 |
| rope_theta | 10000 |
| routed_scaling_factor | 16 |
| scoring_func | softmax |
| seq_aux | true |
| tie_word_embeddings | false |
| topk_group | 3 |
| topk_method | group_limited_greedy |
| torch_dtype | bfloat16 |
| transformers_version | 4.39.3 |
| use_cache | true |
| v_head_dim | 128 |
| vocab_size | 102400 |