Active Parameters
671B
Context Length
128K
Modality
Text
Architecture
Mixture of Experts
KV Cache Rate
13 KB/tok
License
MIT
Release Date
10 Jan 2026
Knowledge Cutoff
May 2025
API Pricing (per 1M)
Input: $0.26 · Output: $0.38
Precision
Context
Workload Memory: 1321.36 GB
FP16 · 8K context
Base Weights: 1318.75 GB
KV Cache: 0.10 GB
Runtime & Graphs: 2.50 GB
16x NVIDIA H200 141GB
1321.4 / 2030.4 GB usable · TP 8 · PP 2
Determine requirements for your specific workload:
Upgrade to Pro
Checkpoint format
FP16
CUDA
>= 11.8
PyTorch
>= 2.1
Transformers
>= 4.46.3
| Architecture | Precision | Attention kernel |
|---|---|---|
Blackwell, Hopper | Native FP16 | FlashMLA |
Ada Lovelace, Ampere | Native FP16 | FlashAttention-2 |
Rank
#100
| Benchmark | Score | Rank |
|---|---|---|
Professional Knowledge | 0.85 | 13 |
Software Engineering | 0.70 | 15 |
Graduate-Level QA | 0.824 | 56 |
Web Development | high 1362 Standard 1326 | 97 109 |
General Text | auto 1423 Standard 1425 | 100 97 |
Coding Index | 0.44 | 99 |
Intelligence Index | auto 0.21 Standard 0.16 | 154 207 |
Coding Archived | 0.74 | 5 |
Overall Rank
#100
Coding Rank
#98
DeepSeek-V3.2 is an advanced 671B Mixture-of-Experts model from DeepSeek optimized for autonomous agents and complex reasoning. Featuring DeepSeek Sparse Attention and integrated thinking modes, it supports long-horizon execution across a 164K context window.
Attention
Attention Structure
DeepSeek Sparse Attention
Attention Heads
128
Key-Value Heads
1
Position Embedding
Absolute Position Embedding
RoPE Theta
10,000
Normalization
RMS Normalization
Activation Function
SwigLU
Dimensions
Base Precision
FP16
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
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.