DeepSeek MoE
Towards Ultimate Expert Specialization in Mixture-of-Experts Language
...The repository introduces fine-grained expert segmentation and shared expert isolation to improve specialization while controlling compute cost. For example, their MoE variant with 16.4B parameters claims comparable or better performance to standard dense models like DeepSeek 7B or LLaMA2 7B using about 40% of the total compute. The repo publishes both Base and Chat variants of the 16B MoE model (deepseek-moe-16b) and provides evaluation results across benchmarks. It also includes a quick start with inference instructions (using Hugging Face Transformers) and guidance on fine-tuning (DeepSpeed, hyperparameters, quantization). The licensing is MIT for code, with a “Model License” applied to the models.