DeepSeek · 2025-09 · Sparse MoE
DeepSeek-V3.2
DeepSeek-V3.2 is a multi-head latent attention (MLA) transformer released by DeepSeek in 2025-09, with 61 layers, hidden size 7168 and a context window of 163,840 tokens.
Layer stack
Attention ×61 · Layers 61
Key facts
| Family | DeepSeek |
|---|---|
| Released | 2025-09 |
| Params | 671 B |
| Active | 37 B |
| Context | 163,840 tokens |
| Attention | MLA · kv_lora_rank 512 |
| Layers | 61 |
| Hidden | 7,168 |
| Heads | 128 |
| Vocab | 129,280 |
| Position | RoPE |
| Norm | RMSNorm |
| Activation | silu |
| Dtype | fp8 |
| Architecture class | DeepseekV32ForCausalLM |
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 37B parameters are active per token.
The configuration ships 1 extra multi-token-prediction (MTP) layers used to accelerate decoding.
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-V3.1
| Model | DeepSeek-V3.2 | DeepSeek-V3.1 | Ratio |
|---|---|---|---|
| model_type | deepseek_v32 | deepseek_v3 | ≠ |
| architectures | DeepseekV32ForCausalLM | DeepseekV3ForCausalLM | ≠ |
| hidden_size | 7168 | 7168 | ≈1 |
| num_hidden_layers | 61 | 61 | ≈1 |
| num_attention_heads | 128 | 128 | ≈1 |
| num_key_value_heads | 128 | 128 | ≈1 |
| intermediate_size | 18432 | 18432 | ≈1 |
| hidden_act | silu | silu | ＝ |
| n_routed_experts | 256 | 256 | ≈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 | 10000 | 10000 | ≈1 |
| max_position_embeddings | 163840 | 163840 | ≈1 |
| rms_norm_eps | 0.000001 | 0.000001 | ≈1 |
| vocab_size | 129280 | 129280 | ≈1 |
| torch_dtype | bfloat16 | bfloat16 | ＝ |
| num_nextn_predict_layers | 1 | 1 | ≈1 |
| attention_bias | false | false | ＝ |
| attention_dropout | 0 | 0 | ＝ |
| bos_token_id | 0 | 0 | ＝ |
| eos_token_id | 1 | 1 | ≈1 |
| ep_size | 1 | 1 | ≈1 |
Most similar architectures
- DeepSeek-V3 (DeepSeek, 2024-12) 1.000
- DeepSeek-V3 Base (DeepSeek, 2024-12) 1.000
- DeepSeek-R1 (DeepSeek, 2025-01) 1.000
- DeepSeek-V3.1 (DeepSeek, 2025-08) 1.000
- Kimi K2 0905 (Kimi, 2025-09) 1.000
- Kimi K2.5 (Kimi, 2026-02) 1.000
Similarity values range from 0 (no shared categorical features) to 1 (identical profiles).
Raw config fields
43 fields
| architectures | DeepseekV32ForCausalLM |
|---|---|
| attention_bias | false |
| attention_dropout | 0 |
| bos_token_id | 0 |
| eos_token_id | 1 |
| ep_size | 1 |
| first_k_dense_replace | 3 |
| hidden_act | silu |
| hidden_size | 7168 |
| index_head_dim | 128 |
| index_n_heads | 64 |
| index_topk | 2048 |
| initializer_range | 0.02 |
| intermediate_size | 18432 |
| kv_lora_rank | 512 |
| max_position_embeddings | 163840 |
| model_type | deepseek_v32 |
| moe_intermediate_size | 2048 |
| moe_layer_freq | 1 |
| n_group | 8 |
| n_routed_experts | 256 |
| n_shared_experts | 1 |
| norm_topk_prob | true |
| num_attention_heads | 128 |
| num_experts_per_tok | 8 |
| num_hidden_layers | 61 |
| num_key_value_heads | 128 |
| num_nextn_predict_layers | 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 | 2.5 |
| scoring_func | sigmoid |
| tie_word_embeddings | false |
| topk_group | 4 |
| topk_method | noaux_tc |
| torch_dtype | bfloat16 |
| transformers_version | 4.44.2 |
| use_cache | true |
| v_head_dim | 128 |
| vocab_size | 129280 |