DeepSeek · 2026-04 · Sparse MoE
DeepSeek-V4-Pro
DeepSeek-V4-Pro is a multi-query attention (MQA) transformer released by DeepSeek in 2026-04, with 61 layers, hidden size 7168 and a context window of 1,048,576 tokens.
Layer stack
Attention ×61 · Layers 61
Key facts
| Family | DeepSeek |
|---|---|
| Released | 2026-04 |
| Params | 1,606.6 B |
| Active | 49 B |
| Context | 1,048,576 tokens |
| Attention | MQA (128:1) |
| Layers | 61 |
| Hidden | 7,168 |
| Heads | 128 |
| Vocab | 129,280 |
| Position | RoPE |
| Norm | RMSNorm |
| Activation | silu |
| Dtype | fp8+fp4 |
| Architecture class | DeepseekV4ForCausalLM |
Architecture overview
Attention
multi-query attention (MQA) — 128 q-heads / 1 kv-heads. RoPE θ=10,000.
Feed-forward / MoE
It is a sparse mixture-of-experts with 384 routed experts and top-6 routing; about 49B 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-V4-Flash Base
| Model | DeepSeek-V4-Pro | DeepSeek-V4-Flash Base | Ratio |
|---|---|---|---|
| model_type | deepseek_v4 | deepseek_v4 | ＝ |
| architectures | DeepseekV4ForCausalLM | DeepseekV4ForCausalLM | ＝ |
| hidden_size | 7168 | 4096 | ×1.75 |
| num_hidden_layers | 61 | 43 | ×1.42 |
| num_attention_heads | 128 | 64 | ×2.00 |
| num_key_value_heads | 1 | 1 | ≈1 |
| head_dim | 512 | 512 | ≈1 |
| hidden_act | silu | silu | ＝ |
| n_routed_experts | 384 | 256 | ×1.50 |
| num_experts_per_tok | 6 | 6 | ≈1 |
| n_shared_experts | 1 | 1 | ≈1 |
| moe_intermediate_size | 3072 | 2048 | ×1.50 |
| qk_rope_head_dim | 64 | 64 | ≈1 |
| q_lora_rank | 1536 | 1024 | ×1.50 |
| rope_theta | 10000 | 10000 | ≈1 |
| max_position_embeddings | 1048576 | 1048576 | ≈1 |
| sliding_window | 128 | 128 | ≈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 |
| expert_dtype | fp4 | fp8 | ≠ |
Most similar architectures
- DeepSeek-V4-Flash (DeepSeek, 2026-04) 0.994
- DeepSeek-V4-Flash Base (DeepSeek, 2026-04) 0.994
- Qwen3.8 2.4T A95B (Qwen, 2026-08) 0.870
- DeepSeek-V2 (DeepSeek, 2024-05) 0.860
- GLM-4.6 (GLM, 2025-10) 0.860
- GLM-4.5 (GLM, 2025-07) 0.859
Similarity values range from 0 (no shared categorical features) to 1 (identical profiles).
Raw config fields
46 fields
| architectures | DeepseekV4ForCausalLM |
|---|---|
| attention_bias | false |
| attention_dropout | 0 |
| bos_token_id | 0 |
| eos_token_id | 1 |
| expert_dtype | fp4 |
| hc_eps | 0.000001 |
| hc_mult | 4 |
| hc_sinkhorn_iters | 20 |
| head_dim | 512 |
| hidden_act | silu |
| hidden_size | 7168 |
| index_head_dim | 128 |
| index_n_heads | 64 |
| index_topk | 1024 |
| initializer_range | 0.02 |
| max_position_embeddings | 1048576 |
| model_type | deepseek_v4 |
| moe_intermediate_size | 3072 |
| n_routed_experts | 384 |
| n_shared_experts | 1 |
| norm_topk_prob | true |
| num_attention_heads | 128 |
| num_experts_per_tok | 6 |
| num_hidden_layers | 61 |
| num_hash_layers | 3 |
| num_key_value_heads | 1 |
| num_nextn_predict_layers | 1 |
| o_groups | 16 |
| o_lora_rank | 1024 |
| q_lora_rank | 1536 |
| qk_rope_head_dim | 64 |
| rms_norm_eps | 0.000001 |
| rope_theta | 10000 |
| routed_scaling_factor | 2.5 |
| scoring_func | sqrtsoftplus |
| sliding_window | 128 |
| swiglu_limit | 10 |
| tie_word_embeddings | false |
| topk_method | noaux_tc |
| torch_dtype | bfloat16 |
| transformers_version | 4.57.1 |
| use_cache | true |
| vocab_size | 129280 |
| compress_rope_theta | 160000 |
| compress_ratios | 128×31 + 4×30 + 0×1 |