DeepSeek-V4-Pro
DeepSeek · DeepSeek V4 · open weights
- Sliding window
- Sparse attention
- Compressed KV
- RoPE
- Partial RoPE
- Pre-norm
- MoE
- Shared expert
- MTP
Facts and where they come from
| Released | 2026-04 | config.jsonconfig.jsonHugging Face repository creation date (api.createdAt) |
|---|---|---|
| Licence | mit | config.jsonconfig.jsonREADME metadata: license |
| Total parameters | 1.6T | labmodel cardREADME: DeepSeek-V4-Pro with 1.6T parameters (49B activated) |
| Active parameters | 49B | labmodel cardREADME: 49B activated |
| Context length | 1M tokens | config.jsonconfig.jsonmax_position_embeddings |
| Norm placement | pre | codemodelling codetransformers 5.18.0 / repo modelling code for deepseek_v4: input_layernorm before attention, post_attention_layernorm before the MLP |
| Norm type | RMSNorm | codemodelling codetransformers 5.18.0 / repo modelling code for deepseek_v4: input_layernorm before attention, post_attention_layernorm before the MLP |
| QK-norm | no | codemodelling codeno q/k normalisation in the attention block |
| Positional encoding | RoPE on 12.5% of each head | config.jsonconfig.jsonqk_rope_head_dim / head_dim |
| Parallel attention and MLP | no | codemodelling codetransformers 5.18.0 / repo modelling code for deepseek_v4: input_layernorm before attention, post_attention_layernorm before the MLP |
Architecture, drawn from the data
compressed 128h, window 128. Each column is one layer: its token mixer above, its feed-forward block below. Paler columns reuse another layer’s keys and values.
Modelled costs
From the cost model, batch size 1. Totals the lab states are in the table above; differences come from rounding, from what a lab counts, or from parts the model does not describe (listed on the about page).
| Parameters (modelled) | 1.57T |
|---|---|
| Active per token (modelled) | 49.4B |
| Without embeddings and output head | 1.57T total, 47.5B active |
| Multi-token-prediction layers (extra) | 25.9B |
| Published weights (Hugging Face count) | 1.6T (packed low-bit tensors, so not comparable) |
| KV cache per token, BF16 (layers that grow with context) | 9.62 KiB |
| KV cache + state at 1M tokens, BF16 | 9.62 GiB |
| Decode FLOPs per token at 4K context | 109 GFLOP |
| Prefill FLOPs for a 4K prompt | 419 TFLOP |
KV cache against context
Every architecture field
| Field | Value | Source |
|---|---|---|
| d_model | 7,168 | config.jsonconfig.jsonhidden_size |
| vocab | 129,280 | config.jsonconfig.jsonvocab_size |
| tied_embeddings | false | config.jsonconfig.jsontie_word_embeddings |
| mixers.csa.type | csa | codemodelling codeDeepSeek V4 compressed attention (CSA/HCA) with a sliding window |
| mixers.csa.heads | 128 | config.jsonconfig.jsonnum_attention_heads |
| mixers.csa.head_dim | 512 | config.jsonconfig.jsonhead_dim |
| mixers.csa.q_lora_rank | 1,536 | config.jsonconfig.jsonq_lora_rank |
| mixers.csa.o_lora_rank | 1,024 | config.jsonconfig.jsono_lora_rank |
| mixers.csa.o_groups | 16 | config.jsonconfig.jsono_groups |
| mixers.csa.window | 128 | config.jsonconfig.jsonsliding_window |
| mixers.csa.indexer.heads | 64 | config.jsonconfig.jsonindex_n_heads |
| mixers.csa.indexer.head_dim | 128 | config.jsonconfig.jsonindex_head_dim |
| mixers.csa.indexer.topk | 1,024 | config.jsonconfig.jsonindex_topk |
| ffns.moe.type | moe | codemodelling codeMoE block |
| ffns.moe.experts | 384 | config.jsonconfig.jsonn_routed_experts |
| ffns.moe.active | 6 | config.jsonconfig.jsonnum_experts_per_tok |
| ffns.moe.d_expert | 3,072 | config.jsonconfig.jsonmoe_intermediate_size |
| ffns.moe.gated | true | codemodelling codeexperts are gated MLPs |
| ffns.moe.shared | 1 | config.jsonconfig.jsonn_shared_experts |
| ffns.moe.d_shared | 3,072 | config.jsonconfig.jsonmoe_intermediate_size |
| layout | 2× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) · 1× csa/moe (ratio=128,indexer=false) · 1× csa/moe (ratio=4,indexer=true) | config.jsonconfig.jsoncompress_ratios (indexer on ratio-4 layers) |
| mtp_layers | 1 | config.jsonconfig.jsonnum_nextn_predict_layers |
Sources
- config.json @ b5968e9
- model card
- modelling code · transformers 5.18.0 modelling code, or the model repository's own modelling file at the pinned revision
Listed in the LLM Architecture Gallery checklist as “DeepSeek V4-Pro (1.6T)” (name only; see about).