aiwals

deepseek-ai/DeepSeek-V3

DeepSeek-V3, a strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token. To achieve efficient inference and cost-effective training, DeepSeek-V3 adopts Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were thoroughly validated in DeepSeek-V2.

  • Provider: deepseek-ai
  • Max tokens: 163840
  • Capabilities: chat, tools, structured output

All models