deepseek-ai/deepseek-r1 is the official distribution project repository that publishes the reasoning-focused DeepSeek-R1 and distilled language models, and its latest release tag is v1.0.0. However, the changes included in this tag are not recorded as release notes in the given record, so they are hard to confirm. What can be covered is the model composition and training approach published in the repository.
The core of this repository is reasoning models, that is, language models that also show the process of solving a problem before giving an answer. According to the README, DeepSeek-R1-Zero was trained with large-scale reinforcement learning (RL) without prior SFT (supervised fine-tuning), while DeepSeek-R1 additionally used cold-start data, that is, sample materials serving as the starting point of training, alongside that reinforcement learning. Along with the two models, dense models based on Llama and Qwen are also released. Distillation is a method of moving learned results into a smaller model. The README explains that DeepSeek-R1 showed performance similar to OpenAI-o1 on mathematics, code, and reasoning tasks. The license is MIT, the star count is 91909, and the repository was created on 2025-01-20.
However, the performance comparison is only a description in the README; no numerical table or independent verification results are presented. The languages used are not recorded in the record, and the only filled-in evaluation item is a momentum score of 3.7 out of 10. The training approach is worth referring to, but from the record alone it is too early to judge what this tag improved and how large the performance difference is. It is better to check the release notes and official benchmarks before using it.