DeepSeek, a Chinese AI startup, has unveiled DeepSeek-R1. This open-source reasoning model matches the performance of OpenAI’s o1 across various tasks, including math, coding, and reasoning, but at a fraction of the cost.
Built on the DeepSeek V3 mixture-of-experts model, R1 represents a significant advancement in open-source AI, narrowing the gap with proprietary models in the race toward artificial general intelligence.
The model leverages pure reinforcement learning and supervised fine-tuning to tackle complex reasoning tasks effectively.
DeepSeek-R1 achieved impressive benchmarks, scoring 79.8% on AIME 2024, 97.3% on MATH-500, and a 2,029 rating on Codeforces, outperforming most human programmers.
While it trails OpenAI’s o1 slightly in general knowledge accuracy, with 90.8% versus 91.8%, its affordability sets it apart.
OpenAI’s o1 costs $15 per million input tokens and $60 per million output tokens, while DeepSeek-R1 operates at just $0.55 and $2.19, respectively, making advanced AI more accessible.
The model was developed using an enhanced version of DeepSeek-R1-Zero, which was trained solely through reinforcement learning, showcasing self-evolution and reasoning abilities.
Addressing issues like readability and language inconsistencies, the researchers combined supervised learning and reasoning-oriented reinforcement learning in a multi-stage process. This approach refined the model further, resulting in DeepSeek-R1’s enhanced capabilities.
DeepSeek-R1’s performance demonstrates the potential of open-source AI to compete with industry leaders. The model is available for public use on the DeepSeek chat platform, similar to ChatGPT, and can be accessed via Hugging Face under an MIT license.
Users can integrate the model through APIs or explore its code and weights. As DeepSeek continues to innovate, R1 signifies a milestone in making high-performance AI affordable and widely available.
