| Property | Value |
|---|---|
| Total Parameters | 236B |
| Active Parameters | 21B per token |
| Context Length | 128k tokens |
| License | DeepSeek Model License |
| Paper | arXiv:2405.04434 |
What is DeepSeek-V2?
DeepSeek-V2 represents a significant advancement in Mixture-of-Experts (MoE) language models, combining economic efficiency with powerful performance. Trained on 8.1 trillion tokens, this model introduces innovative architectures including Multi-head Latent Attention (MLA) and DeepSeekMoE, achieving superior results while reducing training costs by 42.5% and KV cache by 93.3% compared to its predecessors.
Implementation Details
The model employs a sophisticated architecture featuring MLA for attention mechanisms and DeepSeekMoE for Feed-Forward Networks. This design enables efficient inference while maintaining high performance across various tasks.
- BF16 precision format
- Requires 80GB*8 GPUs for inference
- Supports both completion and chat interfaces
- Compatible with Hugging Face Transformers and vLLM
Core Capabilities
- Strong performance on MMLU (78.5%) and BBH (78.9%)
- Exceptional Chinese language understanding (C-Eval: 81.7%, CMMLU: 84.0%)
- Robust coding capabilities (HumanEval: 48.8%, MBPP: 66.6%)
- Advanced mathematical reasoning (GSM8K: 79.2%, Math: 43.6%)
Frequently Asked Questions
Q: What makes this model unique?
DeepSeek-V2's uniqueness lies in its efficient MoE architecture that activates only 21B parameters per token while maintaining the power of a 236B parameter model, offering an optimal balance between performance and resource utilization.
Q: What are the recommended use cases?
The model excels in various applications including general text generation, code development, mathematical problem-solving, and multilingual tasks, with particular strength in Chinese language processing.