What it does
The repository documents DeepSeek-Coder-V2, a code language model built on a Mixture-of-Experts design. Training continued from a mid-stage DeepSeek-V2 checkpoint with 6 trillion more tokens. Compared with the earlier DeepSeek-Coder, the maintainers say the new model covers 338 programming languages instead of 86 and handles 128K tokens of context instead of 16K. Two sizes are released, each as a base and an instruct model: a Lite version with 16B total and 2.4B active parameters, and a full version with 236B total and 21B active parameters. Weights are linked from Hugging Face. The page reports benchmark tables for code generation, completion, fixing, mathematics and general language tasks, comparing the models with several closed and open ones. There are three ways to use it. The DeepSeek website offers a chat interface, DeepSeek Platform provides an OpenAI-compatible API billed pay-as-you-go, and the models can be run locally with Hugging Face Transformers, SGLang or vLLM. The page notes that running the full model in BF16 takes eight 80GB GPUs. The repository code is MIT licensed, the models fall under a separate Model License, and the page says commercial use is supported.
Capabilities
- Mixture-of-Experts code model in 16B and 236B parameter sizes
- Base and instruct versions of each size
- 128K context length
- Support for 338 programming languages
- Code completion, code insertion and chat completion examples
- Hosted chat site and OpenAI-compatible API, or local inference with Transformers, SGLang or vLLM
When to use it
- You want an open-weights model for code generation, completion and fixing
- You want to self-host a coding model and your hardware can support it
- You prefer a hosted API with pay-as-you-go billing over running weights yourself
- You need a model whose licence states that commercial use is supported
