DeepSeek V2 by deepseek-ai

 ยป  All LLMs  ยป  deepseek-ai  ยป  DeepSeek V2   URL Share it on

  Arxiv:2311.18743   Arxiv:2405.04434   Autotrain compatible   Conversational   Custom code   Deepseek v2   Endpoints compatible   Region:us   Safetensors   Sharded   Tensorflow
Model Card on HF ๐Ÿค—: https://huggingface.co/deepseek-ai/DeepSeek-V2 

DeepSeek V2 Benchmarks

nn.n% — How the model compares to the reference models: Anthropic Sonnet 3.5 ("so35"), GPT-4o ("gpt4o") or GPT-4 ("gpt4").
DeepSeek V2 (deepseek-ai/DeepSeek-V2)

DeepSeek V2 Parameters and Internals

Model Type 
Mixture-of-Experts, Language Model
Use Cases 
Areas:
Research, Commercial Applications, Chatbots
Applications:
Language Understanding, Code Generation, Translation, Economical AI Applications
Primary Use Cases:
Text Generation, Conversation AI, Code Classification
Limitations:
Reduced performance on low resource languages and contexts, Complex and high resource computations
Considerations:
Utilize on recommended hardware for efficiency.
Additional Notes 
The model focuses on efficiency with a large parameter architecture for high performance.
Supported Languages 
English (Advanced), Chinese (Advanced), Code (Intermediate)
Training Details 
Data Sources:
High-quality corpus
Data Volume:
8.1 trillion tokens
Methodology:
Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL)
Context Length:
128000
Model Architecture:
Multi-head Latent Attention and DeepSeekMoE architecture
Safety Evaluation 
Methodologies:
Benchmarks, Comparison Tests, Open-ended Evaluation
Findings:
Effective performance on both language and coding benchmarks
Risk Categories:
Fairness, Bias, Misinformation
Ethical Considerations:
Contains Responsible AI guidelines
Responsible Ai Considerations 
Fairness:
Benchmarks evaluate across languages and use cases.
Transparency:
Performance and architecture details publicly shared.
Accountability:
DeepSeek-AI is accountable for model's outputs.
Mitigation Strategies:
Regular updates and evaluation on fairness and bias.
Input Output 
Input Format:
Supports text input, prompts for chat
Accepted Modalities:
text
Output Format:
Generated text with coherent structure
Performance Tips:
Use multipliers and optimization libraries for GPU.
Release Notes 
Version:
V2
Date:
2024
Notes:
Introduction of Mixture-of-Experts, Enhanced efficiency and reduced training costs.
LLM NameDeepSeek V2
Repository ๐Ÿค—https://huggingface.co/deepseek-ai/DeepSeek-V2 
Model Size235.7b
Required VRAM378.4 GB
Updated2025-01-20
Maintainerdeepseek-ai
Model Typedeepseek_v2
Model Files  8.6 GB: 1-of-55   8.6 GB: 2-of-55   8.6 GB: 3-of-55   8.6 GB: 4-of-55   8.6 GB: 5-of-55   8.6 GB: 6-of-55   8.6 GB: 7-of-55   8.6 GB: 8-of-55   8.6 GB: 9-of-55   8.6 GB: 10-of-55   8.6 GB: 11-of-55   8.6 GB: 12-of-55   8.6 GB: 13-of-55   8.6 GB: 14-of-55   8.6 GB: 15-of-55   8.6 GB: 16-of-55   8.6 GB: 17-of-55   8.6 GB: 18-of-55   8.6 GB: 19-of-55   8.6 GB: 20-of-55   8.6 GB: 21-of-55   8.6 GB: 22-of-55   8.6 GB: 23-of-55   8.6 GB: 24-of-55   8.6 GB: 25-of-55   8.6 GB: 26-of-55   8.6 GB: 27-of-55   8.6 GB: 28-of-55   8.6 GB: 29-of-55   8.6 GB: 30-of-55   8.6 GB: 31-of-55   8.6 GB: 32-of-55   8.6 GB: 33-of-55   8.6 GB: 34-of-55   8.6 GB: 35-of-55   8.6 GB: 36-of-55   8.6 GB: 37-of-55   8.6 GB: 38-of-55   8.6 GB: 39-of-55   8.6 GB: 40-of-55   8.6 GB: 41-of-55   8.6 GB: 42-of-55   8.6 GB: 43-of-55   8.6 GB: 44-of-55
Model ArchitectureDeepseekV2ForCausalLM
Licenseother
Context Length163840
Model Max Length163840
Transformers Version4.39.3
Vocabulary Size102400
Torch Data Typebfloat16

Best Alternatives to DeepSeek V2

Best Alternatives
Context / RAM
Downloads
Likes
DeepSeek V2.5 1210160K / 378.4 GB485558239
DeepSeek V2.5160K / 378.4 GB4010684
DeepSeek Coder V2 Instruct160K / 387 GB139676529
...eepSeek Coder V2 Instruct 0724160K / 378.4 GB83999
DeepSeek V2 Chat160K / 378.4 GB1349448
DeepSeek V2 Chat 0628160K / 369.8 GB198171
DeepSeek Coder V2 Base160K / 387 GB18057
Note: green Score (e.g. "73.2") means that the model is better than deepseek-ai/DeepSeek-V2.

Rank the DeepSeek V2 Capabilities

๐Ÿ†˜ Have you tried this model? Rate its performance. This feedback would greatly assist ML community in identifying the most suitable model for their needs. Your contribution really does make a difference! ๐ŸŒŸ

Instruction Following and Task Automation  
Factuality and Completeness of Knowledge  
Censorship and Alignment  
Data Analysis and Insight Generation  
Text Generation  
Text Summarization and Feature Extraction  
Code Generation  
Multi-Language Support and Translation  

What open-source LLMs or SLMs are you in search of? 41636 in total.

Our Social Media →  
Original data from HuggingFace, OpenCompass and various public git repos.
Release v20241227