DeepSeek V3
DeepSeekยทdeepseek/deepseek-v3
Open-source 671B Mixture-of-Experts foundation model trained on 14.8 trillion tokens with DualPipe parallelism and Multi-Head Latent Attention.
Open WeightsOpenTool CallingStructured OutputsNatural Language ProcessingArtificial Intelligence
Context
131,072
Output Limit
8,192
Linked Papers
51
Citations
1.8K
Release Date
Dec 26, 2024
Model Specifications
- Lab / Vendor
- DeepSeek
- Model Family
- DeepSeek
- Modality
- text
- Architecture
- DeepSeek MoE with MLA & Multi-Token Prediction (MTP)
- Parameters
- 671B MoE (37B active)
- Context Window
- 131,072
- Max Output
- 8,192
- Input / Output Price
- $0.27 / $1.1 per 1M
- License
- MIT
Research & Evaluation Papers
Academic evaluations, empirical benchmarks, and publications for DeepSeek V3.
Sort