In a significant leap for artificial intelligence, the Chinese AI lab DeepSeek has unveiled its latest open-source reasoning models, DeepSeek-R1 and DeepSeek-R1 Zero. These models aim to democratize access to advanced reasoning capabilities while delivering competitive performance. The company even claims that DeepSeek-R1 matches OpenAI’s o1 in performance.

Based on its previous ultra-large model, DeepSeek V3, DeepSeek asserts that R1 outperforms o1 on benchmarks such as AIME, MATH-500, and SWE-bench verified.

According to verified benchmarks, R1 achieved:

  • 79.8% on the 2024 American Invitational Mathematics Examination (AIME), edging past o1's 79.2%.
  • 97.3% success rate on MATH-500, exceeding o1's 96.4%.
  • 2,029 Codeforces rating, outperforming 96.3% of human programmers, compared to o1’s 96.6%.

In general knowledge testing, R1 scored 90.8% on the Massive Multitask Language Understanding (MMLU) benchmark, trailing closely behind o1’s 91.8%.

It's available on HuggingFace under an MIT license and you can use the model for free even for commercial purposes.

R1 can fact-check itself, which helps it avoid common problems of non-reasoning models and excel as a reasoning model. It boasts 671 billion parameters, a massive scale that correlates with advanced problem-solving abilities.

For broader accessibility, DeepSeek has also released “distilled” versions of R1 based on Llama and Qwen. These distilled models, ranging from 1.5 billion to 70 billion parameters, enable even the smallest version to run on laptops. You can also access these models on HuggingFace. While the full R1 requires high-performance hardware, it is also available via DeepSeek’s API at 90% to 95% lower prices than OpenAI’s o1.

What's different about DeepSeek-R1 is that it uses reinforcement learning along with supervised fine-tuning instead of chain-of-thought like o1, which is why it is cheaper by 90-95%.

Despite its impressive capabilities, R1 has limitations. As a Chinese-developed model, it undergoes regulation by China’s internet authorities to ensure compliance with “core socialist values.” For instance, R1 avoids responding to questions about Tiananmen Square or Taiwan’s autonomy.

However, the release of DeepSeek-R1 still further closes the gap between closed and open-source models in the race to achieve Artificial general intelligence (AGI) despite these challenges.

DeepSeek-R1 represents more than just a technical achievement; it’s a statement about the potential of open-source AI in a landscape often dominated by proprietary systems. By combining cutting-edge performance with accessible deployment options, DeepSeek is empowering developers worldwide to participate in the AI revolution.

As the race for AGI continues, models like DeepSeek-R1 are a reminder that collaboration and openness are powerful forces in shaping the future of technology.