DeepSeek-Coder

DeepSeek-Coder is a family of coding models for developers who want to run a model on their own hardware or build it into a coding assistant. It includes base models for completion and instruction-tuned models for coding requests written in plain language.

• Completion and infilling: Continue existing code or fill a gap using the surrounding code, with support for context from across a project.
• Code generation: Ask instruction-tuned models to produce code from natural-language requests across multiple programming languages.
• Bug fixing and math: DeepSeek-Coder-V2 supports software defect fixing and mathematical reasoning alongside coding tasks.
• Model choice: Choose among different model sizes, including V2 Lite and full-size models in both base and instruction-tuned forms.

The original models train on repository-level code and English and Chinese text. DeepSeek-Coder-V2 builds on DeepSeek-V2 and uses a mixture-of-experts architecture that activates part of the model for each token. Its larger context window accommodates longer code inputs.

You can download the weights through Hugging Face for local use. DeepSeek also offers a hosted chat interface, separate from running the downloaded models. The original DeepSeek-Coder models are free for research and commercial use. The DeepSeek-Coder-V2 repository code uses the MIT license.

The linked repository distributes DeepSeek-Coder-V2 Base and Instruct models, including Lite variants. Code uses MIT, while weights use the DeepSeek Model License. The project states that V2 supports commercial use subject to those model terms.