DeepSeek Coder is a series of code language models, available in sizes ranging from 1B to 33B parameters. These models have been trained on a massive dataset consisting of 2T tokens, predominantly code (87%) with some natural language (13%) in both English and Chinese. The models support project-level code completion and infilling by utilizing a large window size of 16K and an additional fill-in-the-blank task. They demonstrate leading performance across various benchmarks like HumanEval, MultiPL-E, MBPP, DS-1000, and APPS. The 33B model, deepseek-coder-33b-instruct, is particularly fine-tuned on 2B tokens of instruction data.
Examples of using the model include generating code in response to prompts. Users can employ the model for tasks such as writing a quick sort algorithm in Python by using the transformers library in Python to run the model inference.
The code repository for DeepSeek Coder is licensed under the MIT License and supports commercial use, subject to the Model License. More details on the license can be found in the repository. For further inquiries, users are encouraged to contact the DeepSeek team directly via email.