DeepSeek is an artificial intelligence (AI) company founded in May 2023 in Hangzhou, China, and is founded and backed by High-Flyer, a Chinese hedge fund.
Released in January 2025, DeepSeek R1 is one of the large language models (LLMs) developed by DeepSeek, specifically aimed at enabling advanced logical and mathematical reasoning.
background
High-Flyer was founded by three engineers from Zhejiang University during the 2007~2008 financial crisis and has been trading stocks using machine learning. In 2019, he founded High-Flyer AI, which specializes in researching AI algorithms, and by 2021, he had implemented AI in all his trading strategies. In April 2023, High-Flyer announced the establishment of an independent organization dedicated to the study of artificial general intelligence (AGI), which was officially launched as DeepSeek in May.
Release history
DeepSeek Coder
November 2, 2023: Unveiling the first model, the DeepSeek Coder. It was a fully open source model that could be used commercially.
GitHub: DeepSeek Coder
License: MIT
https://github.com/deepseek-ai/DeepSeek-Coder
DeepSeek LLM
November 29, 2023: Announced DeepSeek LLM, a large language model with 6.7 billion parameters. While it aimed for performance comparable to GPT-4, it faced challenges in computational efficiency and scalability. At the same time, the chat version “DeepSeek Chat” was also released (now upgraded to DeepSeek-V3).
GitHub: DeepSeek-LLM
License: MIT
https://github.com/deepseek-ai/DeepSeek-LLM
DeepSeek-V2
May 2024: DeepSeek-V2 released. It is ranked 7th in the University of Waterloo’s Tiger Lab ranking.
GitHub：DeepSeek-V2
License: MIT
https://github.com/deepseek-ai/DeepSeek-V2
Hugging Face：
https://huggingface.co/deepseek-ai/DeepSeek-V2
DeepSeek-V3
December 2024: Announced “DeepSeek-V3” with 671 billion parameters. It was trained in about 55 days and cost about $5.58 million. It surpassed Llama 3.1 and Qwen 2.5, and showed performance on par with GPT-4o and Claude 3.5 Sonnet.
GitHub：DeepSeek-V3
License: MIT
https://github.com/deepseek-ai/DeepSeek-V3
DeepSeek API Price
https://api-docs.deepseek.com/quick_start/pricing/
DeepSeek-R1
January 2025: DeepSeek-R1 and DeepSeek-R1-Zero are released. Both are Mixture of Experts (MoE) models based on V3-Base with 671 billion total parameters and 37 billion activation parameters. R1-Zero is trained on reinforcement learning only, with no supervised fine-tuning.
It became a hot topic because it has the same performance as OpenAI-o1 and is licensed by MIT.
GitHub: DeepSeek-R1
License: MIT
https://github.com/deepseek-ai/DeepSeek-R1
Download:
Hugging Face Repository: DeepSeek-R1
https://huggingface.co/deepseek-ai/DeepSeek-R1
Hugging Face Repository: DeepSeek-R1-Zero
https://huggingface.co/deepseek-ai/DeepSeek-R1-Zero
Hugging Face Repository: DeepSeek-R1
https://huggingface.co/collections/deepseek-ai/deepseek-r1-678e1e131c0169c0bc89728d
References
DeepSeek official website:
https://www.deepseek.com
Comments