DeepSeek · 2024-01 · Sparse MoE
DeepSeekMoE 16B
DeepSeekMoE 16B is a multi-head attention (MHA) transformer released by DeepSeek in 2024-01, with 28 layers, hidden size 2048 and a context window of 4,096 tokens.
Layer stack
Attention ×28 · Layers 28
Key facts
| Family | DeepSeek |
|---|---|
| Released | 2024-01 |
| Params | 16.2 B |
| Active | 2.6 B |
| Context | 4,096 tokens |
| Attention | MHA (16:16) |
| Layers | 28 |
| Hidden | 2,048 |
| Heads | 16 |
| Vocab | 102,400 |
| Position | RoPE |
| Norm | RMSNorm |
| Activation | silu |
| Dtype | bf16 |
| Architecture class | DeepseekForCausalLM |
Architecture overview
Attention
multi-head attention (MHA) — 16 q-heads / 16 kv-heads. RoPE θ=10,000.
Feed-forward / MoE
It is a sparse mixture-of-experts with 64 routed experts and top-6 routing; about 2.6B 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 DeepSeek LLM 7B
| Model | DeepSeekMoE 16B | DeepSeek LLM 7B | Ratio |
|---|---|---|---|
| model_type | deepseek | llama | ≠ |
| architectures | DeepseekForCausalLM | LlamaForCausalLM | ≠ |
| hidden_size | 2048 | 4096 | ×0.500 |
| num_hidden_layers | 28 | 30 | ×0.933 |
| num_attention_heads | 16 | 32 | ×0.500 |
| num_key_value_heads | 16 | 32 | ×0.500 |
| intermediate_size | 10944 | 11008 | ×0.994 |
| hidden_act | silu | silu | ＝ |
| n_routed_experts | 64 | — | |
| num_experts_per_tok | 6 | — | |
| n_shared_experts | 2 | — | |
| moe_intermediate_size | 1408 | — | |
| rope_theta | 10000 | 10000 | ≈1 |
| max_position_embeddings | 4096 | 4096 | ≈1 |
| rms_norm_eps | 0.000001 | 0.000001 | ≈1 |
| vocab_size | 102400 | 102400 | ≈1 |
| torch_dtype | bfloat16 | bfloat16 | ＝ |
| attention_bias | false | — | |
| attention_dropout | 0 | — | |
| bos_token_id | 100000 | 1 | ×100000.00 |
| eos_token_id | 100001 | 2 | ×50000.50 |
| first_k_dense_replace | 1 | — | |
| initializer_range | 0.02 | 0.02 | ≈1 |
| moe_layer_freq | 1 | — | |
| norm_topk_prob | false | — | |
| pretraining_tp | 1 | 1 | ≈1 |
Most similar architectures
- DeepSeek LLM 7B (DeepSeek, 2023-11) 0.874
- DeepSeek-Coder 6.7B (DeepSeek, 2023-11) 0.873
- Qwen1.5 7B (Qwen, 2024-02) 0.872
- Qwen3.6 35B A3B (Qwen, 2026-04) 0.832
- Kimi Linear 48B A3B (Kimi, 2025-10) 0.831
- GPT-OSS 120B (OpenAI, 2025-08) 0.831
Similarity values range from 0 (no shared categorical features) to 1 (identical profiles).
Raw config fields
30 fields
| architectures | DeepseekForCausalLM |
|---|---|
| attention_bias | false |
| attention_dropout | 0 |
| bos_token_id | 100000 |
| eos_token_id | 100001 |
| first_k_dense_replace | 1 |
| hidden_act | silu |
| hidden_size | 2048 |
| initializer_range | 0.02 |
| intermediate_size | 10944 |
| max_position_embeddings | 4096 |
| model_type | deepseek |
| moe_intermediate_size | 1408 |
| moe_layer_freq | 1 |
| n_routed_experts | 64 |
| n_shared_experts | 2 |
| norm_topk_prob | false |
| num_attention_heads | 16 |
| num_experts_per_tok | 6 |
| num_hidden_layers | 28 |
| num_key_value_heads | 16 |
| pretraining_tp | 1 |
| rms_norm_eps | 0.000001 |
| rope_theta | 10000 |
| scoring_func | softmax |
| tie_word_embeddings | false |
| torch_dtype | bfloat16 |
| transformers_version | 4.36.0 |
| use_cache | true |
| vocab_size | 102400 |