DeepSeek-Coder

DeepSeek-Coder is a family of open code language models trained from scratch on two trillion tokens composed of 87 percent code and 13 percent natural language in English and Chinese. The models come in sizes from 1.3B to 33B parameters, were trained on a project-level corpus with dependency-aware file ordering, and use a fill-in-the-middle objective plus a 16K context window to support repository-scale completion and code infilling across more than 80 programming languages. Instruction-tuned variants handle chat-style coding assistance, and the repository reports that the 33B instruct model outperforms GPT-3.5-turbo on HumanEval, with evaluations also covering MBPP, DS-1000, and APPS. The code is MIT licensed and the model weights ship under a separate DeepSeek model license that permits commercial use. Developers and organizations wanting self-hosted code completion or a base for fine-tuned coding assistants are the typical adopters.