DeepSeek Coder

Open-source series of code language models by DeepSeek-AI

Trained from scratch on 2T tokens (87% code, 13% English/Chinese text)

State-of-the-art results on HumanEval, MultiPL-E, MBPP, DS-1000, and APPS

Project-level code completion and repository-aware infilling

16K context window, extended to 128K in DeepSeek-Coder-V2

Supports 86+ languages (up to 338 in V2)

Multiple parameter sizes: 1B, 5.7B, 6.7B, 33B

Mixture-of-Experts DeepSeek-Coder-V2 comparable to GPT-4-Turbo on coding/math tasks

Pre-trained on project-level code corpora for multi-file understanding

Fill-in-the-blank (FIM) capability for in-place code edits

Flexible, scalable architecture suitable for varied deployment scenarios

Backed by DeepSeek-AI’s open-source ecosystem and infrastructure tooling