DeepSeek Coder is a series of open-source code language models developed by DeepSeek AI.
Model Sizes: DeepSeek Coder is available in various sizes, ranging from 1.3B to 33B parameters, with the latest version (V2) offering a 236B parameter model.
Training Data: The models are trained from scratch on 2 trillion tokens, consisting of 87% code and 13% natural language in both English and Chinese.
Programming Language Support: DeepSeek Coder V2 supports 338 programming languages, a significant increase from the 86 languages supported by the previous version.
Context Length: The latest version (V2) has extended the context length from 16K to 128K tokens.
Performance: DeepSeek Coder achieves state-of-the-art performance among open-source code models on multiple benchmarks, including HumanEval, MultiPL-E, MBPP, DS-1000, and APPS. The V2 model even outperforms some closed-source models like GPT-4 Turbo in coding and math tasks.
Specialized Tasks: The models are pre-trained on project-level code corpus and employ a fill-in-the-blank task, making them particularly effective for project-level code completion and infilling.
Open-Source and Commercial Use: DeepSeek Coder is open-source and allows for both research and unrestricted commercial use under a permissive license.
Availability: There are both free and paid versions available. The latest V2 release includes a free 16B parameter model and a paid 236B parameter model.
DeepSeek Coder represents a significant advancement in open-source code language models, offering capabilities that rival or exceed those of some closed-source alternatives, while providing researchers and developers with full access to unleash its potential in various coding and AI applications.