DeepSeek-V3, a powerful Mixture-of-Experts (MoE) language model with a total of 671B parameters and 37B parameters activated for each token. To achieve efficient inference and cost-effective training, DeepSeek-V3 adopts the Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which have been thoroughly verified in DeepSeek-V2. Moreover, DeepSeek-V3 pioneers an auxiliary-loss-free strategy for load balancing and sets a multi-token prediction training objective for better performance. We pre-train DeepSeek-V3 on 14.8 trillion diverse and high-quality tokens, followed by Supervised Fine-Tuning and Reinforcement Learning stages to fully utilize its capabilities. Comprehensive evaluations show that DeepSeek-V3 outperforms other open-source models and achieves performance comparable to leading closed-source models. Despite its excellent performance, the full training of DeepSeek-V3 only requires 2.788M H800 GPU hours. In addition, its training process is remarkably stable. Throughout the entire training process, we didn't encounter any unrecoverable loss spikes or perform any rollbacks.