"Собрал свои посты про DeepSeek воедино в англоязычном блоге. Когда картинки внутри текста, оно, конечно, удобнее. 1️⃣ ""DeepSeek moment"" (https://gonzoml.substack.com/p/deepseek-moment). The high-level overview of what was special with DeepSeek and what models are there. 2️⃣ ""DeepSeek-V3: Technical Details"" (https://gonzoml.substack.com/p/deepseek-v3-technical-details) Technical details about DeepSeek-V3 architecture: Multi-head Latent Attention (MLA), DeepSeekMoE, Multi-Token Prediction (MTP), FP8 Training, Suggestions on Hardware Design, etc 3️⃣ ""DeepSeek-V3: Training"" (https://gonzoml.substack.com/p/deepseek-v3-training) The training procedure of DeepSeek-V3-Base and DeepSeek-V3. Pre-training, Post-training, Group Relative Policy Optimization (GRPO), etc. 4️⃣ ""DeepSeek-R1: Open model with Reasoning"" (https://gonzoml.substack.com/.../deepseek-r1-open-model...) The DeepSeek-R1 reasoning model, similar to OpenAI o1/o3, Google Gemini 2.0 Flash Thinking, or Alibaba Cloud Qwen QwQ. DeepSeek-R1-Zero, DeepSeek-R1, and DeepSeek-R1-Distill models are covered here."