DeepSeek-V4-Flash
Active Parameters
284B
Context Length
1.05M
Modality
Text
Architecture
Mixture of Experts (MoE)
License
MIT
Release Date
24 Apr 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
30x RTX 4090
24GB VRAM
Datacenter
8x NVIDIA A100
80GB VRAM
Apple Silicon
6x Apple M3 Max
128GB VRAM
1,048,576 tokens
Consumer
35x RTX 4090
24GB VRAM
Datacenter
10x NVIDIA A100
80GB VRAM
Apple Silicon
7x Apple M3 Max
128GB VRAM
Architecture Diagram
Evaluation Benchmarks
Rank
#36
| Benchmark | Score | Rank |
|---|---|---|
Professional Knowledge | 0.864 | 6 |
Data Analysis | 0.79 | 11 |
StackUnseen | 0.843 | 12 |
Agent Arena | high 0.02 | 26 |
Reasoning | 0.87 | 29 |
Web Development | high 1581 | 32 |
Graduate-Level QA | 0.874 | 35 |
General | 0.74 | 37 |
LiveBench Average | 0.74 | 37 |
Agentic Index | max 0.41 | 40 |
Coding | 0.75 | 43 |
Agentic Coding | 0.47 | 43 |
Coding Index | max 0.69 high 0.52 | 48 86 |
Mathematics | 0.80 | 51 |
Intelligence Index | max 0.34 high 0.24 Standard 0.19 | 72 128 169 |
General Text | high 1438 Standard 1436 | 79 83 |
Rankings
Overall Rank
#36
Coding Rank
#37
About DeepSeek-V4-Flash
DeepSeek-V4-Flash is DeepSeek's fast, efficient, and economical MoE model in the V4 series, with 284B total parameters and 13B activated per token. Shares the same hybrid CSA+HCA attention architecture and 1M context support as V4-Pro. DeepSeek-V4-Flash-Max achieves comparable reasoning performance to V4-Pro when given a larger thinking budget. Strong on agentic and coding tasks (SWE-Bench Verified 79.0%, Terminal-Bench 2.0 56.9%), with smaller parameter scale enabling faster response times. Supports Non-think, Think High, and Think Max reasoning modes. Available via API as deepseek-v4-flash. Released open-source under MIT license on April 24, 2026.
Technical Specifications
Attention
Attention Structure
DeepSeek Sparse Attention
Attention Heads
64
Key-Value Heads
1
Attention Head Dimension
512
Position Embedding
Absolute Position Embedding
RoPE Theta
10,000
Sliding Window Attention
Yes
Sliding Window Size
128
Sliding Window Ratio
-
Linear Attention
-
Linear Attention Ratio
-
Normalization
RMS Normalization
Activation Function
Swish
Dimensions
Auxiliary Parameters
-
Hidden Dimension Size
4,096
Number of Layers
43
FFN Intermediate Size (Dense)
2,048
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
-
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.