DeepSeek V4.1 Flash
Active Parameters
552B
Context Length
1.05M
Modality
Text
Architecture
Mixture of Experts (MoE)
License
MIT License
Release Date
10 Sept 2026
Knowledge Cutoff
-
API Pricing (per 1M)
Input: $0.30 · Output: $1.20
System Requirements
VRAM requirements for different quantization methods and context sizes
1,024 tokens
Consumer
63x RTX 4090
24GB VRAM
Datacenter
17x NVIDIA A100
80GB VRAM
Apple Silicon
13x Apple M3 Max
128GB VRAM
1,048,576 tokens
Consumer
69x RTX 4090
24GB VRAM
Datacenter
18x NVIDIA A100
80GB VRAM
Apple Silicon
15x Apple M3 Max
128GB VRAM
Architecture Diagram
Evaluation Benchmarks
Rank
#13
| Benchmark | Score | Rank |
|---|---|---|
Agentic Coding | max 0.77 | 🥇 1 |
General | max 0.81 | 7 |
LiveBench Average | max 0.81 | 7 |
Data Analysis | max 0.79 | 14 |
Coding | max 0.80 | 18 |
Mathematics | max 0.93 | 18 |
Agent Arena | max 0.04 | 18 |
Web Development | max 1620 | 20 |
Graduate-Level QA | 0.909 | 21 |
Reasoning | max 0.87 | 28 |
General Text | max 1473 | 35 |
Intelligence Index | max 0.40 Standard 0.25 | 53 128 |
Rankings
Overall Rank
#13
Coding Rank
#22
About DeepSeek V4.1 Flash
DeepSeek-V4.1-Flash is a multimodal Mixture-of-Experts model with 552B total parameters, activating 8B during prefill and 16B during decoding. It utilizes a Causal Encoder-Decoder architecture with SWA Bounded Replay to significantly reduce KV cache memory across long context windows up to one million tokens.
Technical Specifications
Attention
Attention Structure
Compressed Sparse Attention 2
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
Auxiliary Parameters
-
Hidden Dimension Size
5,120
Number of Layers
40
FFN Intermediate Size (Dense)
-
Multi-Token Prediction Heads
3
Tokenizer
Vocabulary Size
129,280
Mixture of Experts
Total Expert Parameters
16.0B
Number of Experts
384
Active Experts
6
Shared Experts
1
FFN Intermediate Size (per Expert)
2,304
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.