DeepSeek V4 Pro (0813)
Active Parameters
1.61T
Context Length
1.05M
Modality
Reasoning
Architecture
Mixture of Experts
KV Cache Rate
122 KB/tok
License
MIT License
Release Date
13 Aug 2026
Knowledge Cutoff
-
API Pricing (per 1M)
Input: $1.32 · Output: $3.96
Hardware Sizing
Precision
Context
Workload Memory: 3168.09 GB
FP16 · 8K context
Base Weights: 3164.63 GB
KV Cache: 0.95 GB
Runtime & Graphs: 2.50 GB
Minimum Required Hardware
Cluster: 32x NVIDIA H200 141GB
3168.1 / 4060.8 GB usable · TP 8 · PP 4
Determine requirements for datacenter hardware:
Upgrade to Pro
Architecture Diagram
Evaluation Benchmarks
Rank
#45
| Benchmark | Score | Rank |
|---|---|---|
Mathematics | 0.95 | 14 |
Data Analysis | 0.79 | 17 |
General | 0.77 | 22 |
LiveBench Average | 0.77 | 23 |
Agentic Coding | 0.55 | 27 |
Web Development | high 1582 | 31 |
Reasoning | 0.86 | 32 |
Agentic Index | max 0.41 | 35 |
Agent Arena | high -0.3 | 38 |
Coding | 0.77 | 39 |
General Text | high 1465 | 46 |
Coding Index | max 0.69 | 48 |
Intelligence Index | max 0.23 Standard 0.13 | 131 221 |
Rankings
Overall Rank
#45
Coding Rank
#40
About DeepSeek V4 Pro (0813)
DeepSeek V4 Pro 0813 represents the general availability release of DeepSeek's flagship mixture-of-experts model. It provides frontier reasoning, complex software engineering capabilities, and extended context analysis.
Technical Specifications
Attention
Attention Structure
DeepSeek Sparse Attention
Attention Heads
128
Key-Value Heads
1
Attention Head Dimension
512
Position Embedding
ROPE
RoPE Theta
10,000
Sliding Window Attention
Yes
Sliding Window Size
128
Normalization
RMS Normalization
Activation Function
SwigLU
Dimensions
Hidden Dimension Size
7,168
Number of Layers
61
Multi-Token Prediction Heads
1
Tokenizer
Vocabulary Size
129,280
Mixture of Experts
Total Expert Parameters
87.8B
Number of Experts
384
Active Experts
6
Shared Experts
1
FFN Intermediate Size (per Expert)
3,072
Dense Layers Before MoE
0
Resources
About DeepSeek V4
DeepSeek-V4 is DeepSeek's latest generation of highly efficient Mixture-of-Experts language models, featuring a novel hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) that dramatically improves long-context efficiency. Pre-trained on 32T+ tokens with a comprehensive post-training pipeline including domain-specific expert cultivation and unified model consolidation. Both V4-Pro and V4-Flash support 1M context length as standard, with three reasoning effort modes (Non-think, Think High, Think Max). Released open-source under MIT license on April 24, 2026.