DeepSeek Core Technologies and Model Innovations: DeepSeek‑V3 and DeepSeek‑R1 Technical Overview

In early 2025 DeepSeek became a focal point in the AI field, with its DeepSeek‑V3 and DeepSeek‑R1 models causing industry‑wide ripples.

DeepSeek‑V3 is a 671‑billion‑parameter mixture‑of‑experts (MoE) model where each token activates 37‑billion parameters; it was pretrained on 14.8 trillion high‑quality tokens using MLA and MoE architectures, achieving strong performance with low cost.

DeepSeek‑R1 builds on V3 as an inference model, employing reinforcement learning (RL) during post‑training to dramatically improve reasoning abilities on mathematics, code, and natural‑language tasks, rivaling OpenAI‑o1.

The V3 architecture is based on a Transformer‑MoE backbone with innovations such as small expert models, multi‑head latent attention, loss‑free load balancing, and the MTP token‑prediction technique, which together boost performance.

Training leverages DeepSeek's lightweight distributed framework HAI‑LLM, overcoming cross‑node MoE communication bottlenecks and being the first open‑source model to use FP8 mixed‑precision training.

For inference, V3 separates prefilling and decoding stages and adopts a redundant‑expert strategy to increase speed while maintaining stability.

DeepSeek‑R1 Technical Breakthroughs

1. Pure reinforcement‑learning training (RL) replaces traditional supervised fine‑tuning (SFT) and RLHF, allowing the model to self‑evolve without any labeled data.

2. The GRPO algorithm improves upon PPO/DPO by grouping N candidate answers, using the group’s average reward as a baseline, eliminating the need for a separate value model and reducing training complexity.

3. Result‑oriented reward models are adopted to avoid reward hacking and reduce the need for extensive annotation.

The training pipeline includes a “cold‑start + multi‑stage RL” strategy: a cold‑start phase fine‑tunes the base model with thousands of high‑quality chain‑of‑thought examples to improve readability, followed by two stages of RL that further refine performance.

These innovations illustrate the principle of letting models explore freely—"Don’t teach, incentivize"—to achieve higher performance as compute scales.

The article also announces the book “DeepSeek Core Technology Revealed”, which systematically analyzes DeepSeek’s architecture, training optimizations, inference tricks, and open‑source projects such as FlashMLA, DeepEP, DeepGEMM, DualPipe, and EPLB, offering readers deep insights into large‑model engineering.

The promotion includes a limited‑time half‑price offer for the full‑color printed book, with the authors Lu Jing and Dai Zhishi—experts in AI research and architecture—providing credibility and further resources for readers interested in large‑model technologies.