DeepSeek: Training Process, Working Principles, and Recent Innovations

DeepSeek has quickly become a leading Chinese AI model, surpassing OpenAI's ChatGPT in performance and topping App Store rankings. Its success stems from a two‑stage training pipeline: pre‑training on massive web‑scale text data followed by post‑training to align model behavior with human expectations.

Pre‑training stage aims to teach the model general language patterns by predicting the next token in billions of sentences sourced from public datasets such as Common Crawl. This phase uses a single loss function, massive compute, and yields a foundational model capable of autoregressive token prediction.

Post‑training stage refines the model through two main methods: instruction tuning (adding specific formats so the model can follow user commands) and reinforcement learning from human feedback (collecting pairwise preference data, using contrastive loss to teach the model to produce higher‑quality, human‑preferred responses).

Model working principles differ between DeepSeek‑V3 and DeepSeek‑R1. V3 is an instruction‑following model that generates concise, markdown‑styled answers directly from prompts. R1 is a reasoning model that first produces a chain‑of‑thought explanation before delivering the final answer, making it better suited for complex logical or mathematical tasks.

Optimization and innovation include exposing intermediate reasoning steps to users, combining inference training with supervised fine‑tuning (SFT), and leveraging strategy‑based evaluation methods such as game‑environment testing to assess model capabilities beyond simple token‑level metrics.

Performance-wise, DeepSeek‑R1 runs with an average latency of about 2 seconds, offering API throughput comparable to GPT‑4.5 and four times faster than GPT‑4, while maintaining high accuracy on reasoning‑heavy queries.

The article concludes by noting the rise of reasoning‑oriented LLMs (e.g., DeepSeek‑R1, DeepSeek‑Reasoning, OpenAI o3‑mini) and encourages readers to follow the author for further technical deep‑dives.