DeepSeek-V3.1
Active Parameters
671B
Context Length
128K
Modality
Text
Architecture
Mixture of Experts
KV Cache Rate
1708 KB/tok
License
MIT License
Release Date
21 Aug 2025
Knowledge Cutoff
-
API Pricing (per 1M)
Input: $1.23 · Output: $4.94
Hardware Sizing
Precision
Context
Workload Memory: 1334.60 GB
FP16 · 8K context
Base Weights: 1318.75 GB
KV Cache: 13.34 GB
Runtime & Graphs: 2.50 GB
Minimum Required Hardware
Cluster: 16x NVIDIA H200 141GB
1334.6 / 2030.4 GB usable · TP 8 · PP 2
Determine requirements for datacenter hardware:
Upgrade to Pro
Architecture Diagram
Evaluation Benchmarks
Rank
#114
| Benchmark | Score | Rank |
|---|---|---|
Professional Knowledge | 0.837 | 19 |
StackUnseen | 0.481 | 23 |
Graduate-Level QA | 0.749 | 78 |
Agentic Index | auto 0.07 | 99 |
Coding Index | 0.43 | 102 |
General Text | auto 1416 Standard 1417 | 106 105 |
Intelligence Index | auto 0.14 Standard 0.14 | 206 204 |
Rankings
Overall Rank
#114
Coding Rank
#97
About DeepSeek-V3.1
A hybrid model that supports both "thinking" and "non-thinking" modes for chat, reasoning, and coding. It's a Mixture-of-Experts (MoE) model with a massive context length and efficient architecture.
Technical Specifications
Attention
Attention Structure
Multi-Head Attention
Attention Heads
128
Key-Value Heads
128
Position Embedding
ROPE
RoPE Theta
10,000
Normalization
RMS Normalization
Activation Function
SwigLU
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
8
Shared Experts
1
FFN Intermediate Size (per Expert)
2,048
Dense Layers Before MoE
3
Resources
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.