DeepSeek AI has released the general-purpose large language model DeepSeek LLM 67B, which has been fully open-sourced. Compared with the same level model LLaMA2 70B, it performs better on nearly 20 public evaluation lists in English and Chinese, especially in reasoning, math, programming, etc. DeepSeek has open-sourced both 7B and 67B scale models, both including the base model and the command fine-tuning model, which are free for commercial use without application. In addition, DeepSeek has also made nine model checkpoints from the midway point of training available for download. In order to further verify the generalization effect of the model outside the real samples, DeepSeek adopts a series of never-before-seen exam questions covering programming, mathematics, instruction following, etc., to evaluate the real ability of the large model in a relatively objective and fair way.
Hugging Face Home:
https://huggingface.co/deepseek-ai
GitHub repositories:
https://github.com/deepseek-ai/DeepSeek-LLM