DeepSeek Coder
About
The DeepSeek Coder family includes a range of open-source code language models specifically designed for handling large codebases. Trained on an expansive dataset of 2 trillion tokens, primarily composed of code (87%) and a mix of English and Chinese natural language data (13%), these models are available in sizes from 1.3 billion to 33 billion parameters. This range gives users the flexibility to choose models that align with their computational resources and specific needs. With pre-training on a high-quality project-level code corpus and using a 16K window size, the models excel in code generation and infill tasks. They demonstrate state-of-the-art performance on various open-source code benchmarks and often outperform some proprietary models. Released under a permissive license, DeepSeek Coder models support both research and commercial applications, offering significant capabilities to developers in coding projects 145.