DeepSeek V4 Flash (0731)
Active Parameters
284B
Context Length
1.31M
Modality
Reasoning
Architecture
Mixture of Experts (MoE)
License
MIT License
Release Date
31 Jul 2026
Knowledge Cutoff
-
API Pricing (per 1M)
Input: $0.44 · Output: $1.32
System Requirements
VRAM requirements for different quantization methods and context sizes
1,024 tokens
Consumer
32x RTX 4090
24GB VRAM
Datacenter
9x NVIDIA A100
80GB VRAM
Apple Silicon
7x Apple M3 Max
128GB VRAM
1,310,720 tokens
Consumer
40x RTX 4090
24GB VRAM
Datacenter
11x NVIDIA A100
80GB VRAM
Apple Silicon
8x Apple M3 Max
128GB VRAM
Architecture Diagram
Evaluation Benchmarks
Rank
#31
| Benchmark | Score | Rank |
|---|---|---|
Data Analysis | 0.79 | 8 |
Reasoning | 0.87 | 23 |
Agent Arena | high 0.02 | 24 |
General | 0.74 | 30 |
LiveBench Average | 0.74 | 30 |
Agentic Index | max 0.42 | 35 |
Coding | 0.75 | 35 |
Agentic Coding | 0.47 | 35 |
Mathematics | 0.87 | 35 |
Coding Index | max 0.69 | 49 |
Intelligence Index | max 0.34 | 51 |
Rankings
Overall Rank
#31
Coding Rank
#34
About DeepSeek V4 Flash (0731)
DeepSeek V4 Flash 0731 is a 284B sparse mixture-of-experts model with 13B active parameters per token from DeepSeek. It features a re-post-trained revision designed for fast coding, reasoning, and multi-turn agent workflows.
Technical Specifications
Attention
Attention Structure
DeepSeek Sparse Attention
Attention Heads
64
Key-Value Heads
1
Attention Head Dimension
512
Position Embedding
ROPE
RoPE Theta
10,000
Sliding Window Attention
Yes
Sliding Window Size
128
Sliding Window Ratio
-
Linear Attention
No
Linear Attention Ratio
-
Normalization
RMS Normalization
Activation Function
SwigLU
Dimensions
Hidden Dimension Size
4,096
Number of Layers
43
FFN Intermediate Size (Dense)
-
Multi-Token Prediction Heads
1
Tokenizer
Vocabulary Size
129,280
Mixture of Experts
Total Expert Parameters
13.0B
Number of Experts
256
Active Experts
6
Shared Experts
1
FFN Intermediate Size (per Expert)
2,048
Dense Layers Before MoE
-
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.