The difference between DeepSeek-V3 and DeepSeek-R1-Zero

[helperfunc]

Can we think the difference between DeepSeek-V3 and DeepSeek-R1-Zero is only with or without supervised fine-tuning during post-training?

[fawern]

Yes, the primary difference between DeepSeek-R1-Zero and DeepSeek-R1 lies in the use of supervised fine-tuning during post-training. DeepSeek-R1-Zero is trained exclusively using reinforcement learning without any supervised fine-tuning. In contrast, DeepSeek-R1 incorporates a supervised fine-tuning phase to enhance readability and coherence in its outputs.

[fawern]

here are some sources:
https://aipapersacademy.com/deepseek-r1/?utm_source=chatgpt.com
https://www.datacamp.com/blog/deepseek-r1?utm_source=chatgpt.com