DeepSeek-V3 671B
Active Parameters
671B
Context Length
131K
Modality
Text
Architecture
Mixture of Experts (MoE)
License
DeepSeek Model License
Release Date
27 Dec 2024
Knowledge Cutoff
Jul 2024
API Pricing (per 1M)
Input: $1.14 · Output: $4.56
System Requirements
VRAM requirements for different quantization methods and context sizes
1,024 tokens
Consumer
80x RTX 4090
24GB VRAM
Datacenter
21x NVIDIA A100
80GB VRAM
Apple Silicon
17x Apple M3 Max
128GB VRAM
131,072 tokens
Consumer
118x RTX 4090
24GB VRAM
Datacenter
29x NVIDIA A100
80GB VRAM
Apple Silicon
25x Apple M3 Max
128GB VRAM
Architecture Diagram
Evaluation Benchmarks
Rank
#144
| Benchmark | Score | Rank |
|---|---|---|
StackUnseen | 0.439 | 27 |
Software Engineering | 0.42 | 31 |
Professional Knowledge | 0.759 | 36 |
Graduate-Level QA | 0.684 | 87 |
General Text | 1395 | 115 |
Agentic Index | 0.8 | 122 |
Coding Index | 0.23 | 126 |
Intelligence Index | 0.10 | 263 |
StackEval Archived | 0.976 | 🥈 2 |
General Knowledge Reference | 0.885 | 5 |
QA Assistant Archived | 0.953 | 9 |
Coding Archived | 0.55 | 11 |
Summarization Archived | 0.806 | 11 |
Rankings
Overall Rank
#144
Coding Rank
#116
About DeepSeek-V3 671B
DeepSeek-V3 is a 671B-parameter Mixture-of-Experts foundation model activating 37B parameters per token for efficient general language processing. Utilizing Multi-Head Latent Attention and Multi-Token Prediction, it excels across coding, math, and 128K long contexts.
Technical Specifications
Attention
Attention Structure
Multi-Layer Attention
Attention Heads
128
Key-Value Heads
128
Position Embedding
ROPE
RoPE Theta
10,000
Normalization
RMS Normalization
Activation Function
Swish
Dimensions
Hidden Dimension Size
7,168
Number of Layers
61
FFN Intermediate Size (Dense)
2,048
Multi-Token Prediction Heads
1
Tokenizer
Vocabulary Size
129,280
Mixture of Experts
Total Expert Parameters
37.0B
Number of Experts
257
Active Experts
9
Shared Experts
1
FFN Intermediate Size (per Expert)
2,048
Dense Layers Before MoE
3
About DeepSeek-V3
DeepSeek-V3 is a Mixture-of-Experts (MoE) language model comprising 671B parameters with 37B activated per token. Its architecture incorporates Multi-head Latent Attention and DeepSeekMoE for efficient inference and training. Innovations include an auxiliary-loss-free load balancing strategy and a multi-token prediction objective, trained on 14.8T tokens.