DeepSeek-Coder
Open-weight coding models with downloadable weights, base and instruction-tuned variants, and support for project-level code completion.
DeepSeek-Coder is a family of code language models for developers who want downloadable models for their own coding tools. It generates code, completes existing code, and fills gaps within a file using the surrounding context. DeepSeek also offers a hosted chat interface; the downloadable weights are separate from that service.
The family includes base models for code completion and instruction-tuned models that respond to natural-language requests. Training on repository-level code helps the models work with context beyond an isolated snippet. The original models also train on English and Chinese text, and come in different sizes for different computing requirements. Weights are available through Hugging Face, and the models are available for research and commercial use under DeepSeek model terms.
DeepSeek-Coder-V2 builds on DeepSeek-V2 and uses a mixture-of-experts architecture, which activates only part of the model for each token. It includes Lite and full-size models, each with base and instruction-tuned variants. Its capabilities cover code generation, completion, bug fixing, and mathematical reasoning. V2 code uses MIT, while its model weights use the separate DeepSeek Model License.
The published evaluations cover Python and Java repository completion, filling missing code, and fixing software defects. They also test generated code and math problems, giving developers task-specific results to compare when choosing a model for a coding assistant.
