Introduction
DeepSeek is a Chinese AI tool, similar to ChatGPT, developed by the startup DeepSeek, founded by Liang Wenfeng. It is widely used for coding suggestions, helping people write blogs, summarizing articles, conducting research across various fields, and much more. It was the most downloaded app on mobile platforms in January 2025.
Evolution of DeepSeek
Different Versions of DeepSeek
DeepSeek Coder
- Designed for Programming activities.
- Trained on a dataset that is 87% code, and 13% on languages such as Chinese and English.
DeepSeek LLM
- Multipurpose language model.
- Built on a dataset of 2 trillion tokens in English and Chinese.
- Has a base and chat model for research and application development.
DeepSeek V2
- Has 236 billion parameters, where 21 billion are activated per token during processing.
- Focused on maximum output with minimal training inputs.
DeepSeek V3
- This model has 671 billion parameters, with a wide range of complex architecture that activates different parts of the model based on the tasks.
DeepSeek-R1
- This model is trained via large-scale reinforcement learning without supervised fine tuning.
- This model focuses on efficiency and cost effectiveness.
Cutting-edge Training Techniques
Reinforced Learning
- DeepSeek uses continuous learning to correct its past mistakes and enhance its current responses.
- With continuous learning, DeepSeek improves dynamically, offering more accurate judgments and refined responses.
Reward Learning
- Reward learning is based on positive reinforcement, enabling the model to generate accurate results across various domains.
Model Distillation
- The Reasoning patterns of large models can be distilled into smaller models, resulting in better performance compared to reasoning patterns through reinforced learning on small models.
- This results in using low computation powers like CPU’s and memory units.