DeepSeek has officially released and open-sourced its latest large language model, R1. This model performs exceptionally well and is considered comparable to OpenAI's official version o1. This initiative not only marks another significant breakthrough in domestic AI technology but also provides new options for global AI developers.

DeepSeek R1 has extensively applied reinforcement learning techniques during the post-training phase, significantly enhancing the model's reasoning capabilities even with very few labeled data. In key tasks such as mathematics, coding, and natural language reasoning, DeepSeek R1's performance is on par with OpenAI's official version o1, demonstrating its powerful capabilities.

image.png

To give back to the open-source community, DeepSeek has also open-sourced two models, DeepSeek-R1 and DeepSeek-R1-Zero, both with a parameter size of 660B. Additionally, DeepSeek has open-sourced six smaller models through model distillation technology, including models with 32B and 70B parameters. These smaller models surpass OpenAI's o1-mini in multiple capabilities, further enriching the open-source ecosystem.

image.png

In terms of API pricing, DeepSeek has also demonstrated an open approach: a cache hit costs only 1 yuan per million input tokens, while a miss costs 4 yuan; output tokens are priced at 16 yuan per million, making the overall pricing more competitive.

Importantly, DeepSeek R1 is licensed under the standard MIT License, allowing users unrestricted commercial use. Additionally, DeepSeek encourages users to utilize the outputs of R1 to train other models, further promoting the popularization and development of AI technology. The open-sourcing of DeepSeek R1 will undoubtedly provide global developers with more powerful tools and inject new vitality into the innovation and application of AI technology, signaling the accelerated arrival of an era of AI technology democratization.

Paper: https://github.com/deepseek-ai/DeepSeek-R1/blob/main/DeepSeek_R1.pdf

API Documentation: https://api-docs.deepseek.com/en/guides/reasoning_model