deepseek-ai/deepseek-coder-v2

DeepSeek-Coder-V2 is an open-source AI model built specifically for writing and understanding code. It was created by DeepSeek AI and trained on an additional 6 trillion tokens of text beyond its predecessor, giving it a strong grasp of programming tasks, math reasoning, and general language. The model supports 338 programming languages and can handle very long inputs, up to 128,000 tokens at a time, which is enough to feed it an entire large codebase at once. The model uses an architecture called Mixture-of-Experts, where only a portion of the model's total parameters are active on any given request. The large version has 236 billion total parameters but activates only 21 billion at inference time, which reduces the compute required to run it. A smaller version with 16 billion total parameters is also available, activating just 2.4 billion at a time. In the benchmark results shown in the README, the large instruct version scores comparably to GPT-4-Turbo on standard code generation and mathematical reasoning tests, and outperforms several other open-source models of similar size. Four model variants are available for download on Hugging Face: a base and an instruct version for each of the two size tiers. To use the model locally, you would load it through a library called Transformers (from Hugging Face) and run it on hardware with enough GPU memory. The README includes code samples showing how to load the model and send it a question. An API is also available for those who do not want to run the model themselves. This repository holds the model documentation, download links, benchmark tables, and usage examples. The model weights themselves are hosted on Hugging Face. The code portions of the repository are released under the MIT license, while the model weights carry a separate model license.