DeepSeek · 2023-11 · Dense
DeepSeek-Coder 6.7B
DeepSeek-Coder 6.7B is a multi-head attention (MHA) transformer released by DeepSeek in 2023-11, with 32 layers, hidden size 4096 and a context window of 16,384 tokens.
Layer stack
Attention ×32 · Layers 32
Key facts
| Family | DeepSeek |
|---|---|
| Released | 2023-11 |
| Params | 6.7 B |
| Context | 16,384 tokens |
| Attention | MHA (32:32) |
| Layers | 32 |
| Hidden | 4,096 |
| Heads | 32 |
| Vocab | 32,256 |
| Position | RoPE |
| Norm | RMSNorm |
| Activation | silu |
| Dtype | bf16 |
| Architecture class | LlamaForCausalLM |
Architecture overview
Attention
multi-head attention (MHA) — 32 q-heads / 32 kv-heads. RoPE θ=100,000.
Feed-forward / MoE
It is a dense model: all 6.7B parameters participate in computing every 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 most-similar Qwen1.5 7B
| Model | DeepSeek-Coder 6.7B | Qwen1.5 7B | Ratio |
|---|---|---|---|
| model_type | llama | qwen2 | ≠ |
| architectures | LlamaForCausalLM | Qwen2ForCausalLM | ≠ |
| hidden_size | 4096 | 4096 | ≈1 |
| num_hidden_layers | 32 | 32 | ≈1 |
| num_attention_heads | 32 | 32 | ≈1 |
| num_key_value_heads | 32 | 32 | ≈1 |
| intermediate_size | 11008 | 11008 | ≈1 |
| hidden_act | silu | silu | ＝ |
| rope_theta | 100000 | 1000000 | ×0.100 |
| max_position_embeddings | 16384 | 32768 | ×0.500 |
| sliding_window | — | 32768 | |
| rms_norm_eps | 0.000001 | 0.000001 | ≈1 |
| vocab_size | 32256 | 151936 | ×0.212 |
| torch_dtype | bfloat16 | bfloat16 | ＝ |
| bos_token_id | 32013 | 151643 | ×0.211 |
| eos_token_id | 32014 | 151643 | ×0.211 |
| initializer_range | 0.02 | 0.02 | ≈1 |
| pretraining_tp | 1 | — | |
| tie_word_embeddings | false | false | ＝ |
| transformers_version | 4.34.1 | 4.37.0 | ≠ |
| use_cache | true | true | ＝ |
Most similar architectures
- Qwen1.5 7B (Qwen, 2024-02) 1.000
- DeepSeek LLM 7B (DeepSeek, 2023-11) 0.999
- LLaMA 1 7B (Meta, 2023-02) 0.884
- DeepSeekMoE 16B (DeepSeek, 2024-01) 0.873
- Hunyuan-TurboS 7B (Hunyuan, 2026-05) 0.789
- Llama 3.1 8B (Meta, 2024-07) 0.789
Similarity values range from 0 (no shared categorical features) to 1 (identical profiles).
Raw config fields
20 fields
| architectures | LlamaForCausalLM |
|---|---|
| bos_token_id | 32013 |
| eos_token_id | 32014 |
| hidden_act | silu |
| hidden_size | 4096 |
| initializer_range | 0.02 |
| intermediate_size | 11008 |
| max_position_embeddings | 16384 |
| model_type | llama |
| num_attention_heads | 32 |
| num_hidden_layers | 32 |
| num_key_value_heads | 32 |
| pretraining_tp | 1 |
| rms_norm_eps | 0.000001 |
| rope_theta | 100000 |
| tie_word_embeddings | false |
| torch_dtype | bfloat16 |
| transformers_version | 4.34.1 |
| use_cache | true |
| vocab_size | 32256 |