Active Parameters
290.9B
Context Length
1.05M
Modality
Text
Architecture
Mixture of Experts
KV Cache Rate
86 KB/tok
License
MIT License
Release Date
31 Aug 2026
Knowledge Cutoff
-
API Pricing (per 1M)
Input: $0.44 · Output: $1.32
Precision
Context
Workload Memory: 583.01 GB
FP16 · 8K context
Base Weights: 579.84 GB
KV Cache: 0.67 GB
Runtime & Graphs: 2.50 GB
8x NVIDIA H200 141GB
583.0 / 1015.2 GB usable · TP 8
Determine requirements for your specific workload:
Upgrade to Pro
Checkpoint format
FP16
CUDA
>= 11.8
PyTorch
>= 2.1
Transformers
>= 5.0.0
| Architecture | Precision | Attention kernel |
|---|---|---|
Blackwell, Hopper | Native FP16 | FlashMLA |
Ada Lovelace, Ampere | Native FP16 | FlashAttention-2 |
Rank
#36
| Benchmark | Score | Rank |
|---|---|---|
Agentic Coding | 0.65 | 5 |
Data Analysis | 0.79 | 12 |
Agentic Index | max 0.47 | 23 |
General | 0.77 | 25 |
LiveBench Average | 0.77 | 26 |
Reasoning | 0.85 | 35 |
Mathematics | 0.88 | 42 |
Coding Index | max 0.65 | 59 |
Coding | 0.68 | 60 |
Intelligence Index | max 0.35 | 67 |
Overall Rank
#36
Coding Rank
#52
DeepSeek-V4-Flash-Vision-Exp is an experimental multimodal vision-language model based on DeepSeek's V4 architecture. It delivers fast visual reasoning, document parsing, and multimodal instruction following.
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
Normalization
RMS Normalization
Activation Function
SwigLU
Dimensions
Base Precision
FP16
Hidden Dimension Size
4,096
Number of Layers
43
Multi-Token Prediction Heads
3
Tokenizer
Vocabulary Size
129,280
Mixture of Experts
Total Expert Parameters
20.4B
Number of Experts
256
Active Experts
6
Shared Experts
1
FFN Intermediate Size (per Expert)
2,048
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.