DeepSeek-V3

DeepSeek-V3 is a Mixture-of-Experts language model with 671B total parameters and 37B activated per token. It uses Multi-head Latent Attention and DeepSeekMoE architectures, was pre-trained on 14.8 trillion tokens, and then underwent supervised fine-tuning and reinforcement learning. The repository provides the model weights and inference code.

GitHub repo Directory import
0 shouts Added Sep 23, 2026 github.com
github.com/deepseek-ai/DeepSeek-V3
DeepSeek-V3 website preview

Links

Tags