unlock-deepseek

Why we included this project

Unlock DeepSeek reads the DeepSeek technical reports one paper at a time and pairs each one with reproduction code, so you can see how the models are built rather than just calling them through an API. The coverage runs from DeepSeek-LLM through the MoE architecture work, the V3 series, and the R1 reasoning model, with chapters on sparse attention, reinforcement learning for reasoning, and training infrastructure. The authors also set DeepSeek next to Kimi, GLM, and MiniMax to show where the different AGI research paths diverge. The intended reader already has basic LLM knowledge and university-level math, which makes it a good fit for researchers, graduate students, and practitioners who want more than a blog summary. Note that the project is still in alpha, with several chapters under construction, so treat it as a work-in-progress study companion rather than a finished reference.