DeepSeek's Large Model Shakes Up the Global AI Community!

Recently, DeepSeek announced the official release of its latest artificial intelligence model, DeepSeek-R1, along with the open-sourcing of the model's weights. DeepSeek-R1 extensively utilized reinforcement learning technology during the post-training phase, significantly enhancing the model's reasoning capabilities with minimal annotated data.

DeepSeek-R1's performance in tasks such as mathematics, coding, and natural language reasoning rivals that of OpenAI's o1 official version. Data shows that DeepSeek-R1's scores in algorithmic coding scenarios (Codeforces) and knowledge-based tests (GPQA, MMLU) are slightly lower than those of OpenAI o1. However, in engineering coding scenarios (SWE-Bench Verified) and the American Mathematics Competitions (AIME 2024, MATH), DeepSeek-R1 outperforms OpenAI o1.

Additionally, it's noteworthy that the DeepSeek-R1 API service is priced at 1 yuan per million input tokens (cache hit) / 4 yuan (cache miss), and 16 yuan per million output tokens, with the output API price being only 3% of OpenAI o1's. Behind the low price is still a show of strength, as the pricing power demonstrates technical prowess—the ability to reduce costs at the AI infrastructure level.