DeepSeek-Coder-V2

Why we included this project

DeepSeek-Coder-V2 is a Mixture-of-Experts model built for code, and it's one of the stronger open-weight options if you want commercial-grade coding performance without depending on a closed API. The repo carries base and instruct variants at 16B and 236B parameters; the larger model scores on HumanEval and similar benchmarks at a level that holds up next to commercial offerings. Because it was continued-pretrained from DeepSeek-V2, it keeps the general language and math reasoning of that model while covering 338 programming languages and a 128K context window. If you self-host inference, fine-tune for internal tooling, or maintain code completion and repair pipelines, the weights, evaluation tables, and usage examples are all here. The 16B Lite variant is a practical choice for smaller deployments that still want solid code generation.