Last updated 07-01-2026
Category:
Reviews:
Join thousands of AI enthusiasts in the World of AI!
DeepSeek Coder
DeepSeek Coder is a family of open-source code language models from DeepSeek. The series includes base and instruct variants from 1B to 33B parameters, each trained from scratch on 2 trillion tokens that are 87% code and 13% natural language in English and Chinese. Weights are published on Hugging Face, and you can try the models through chat.deepseek.com or run them locally with the examples in the GitHub repository.
The models target project-level work, not just single-line snippets. Training uses a 16K token window and a fill-in-the-blank task so the model can complete or insert code across multiple files in a repository. Published benchmark results on the repo show DeepSeek-Coder-Base-33B ahead of CodeLlama-34B on HumanEval Python, HumanEval Multilingual, MBPP, and DS-1000. The instruct-tuned 33B model is reported to beat GPT-3.5-turbo on HumanEval and match it on MBPP.
The repository bundles inference walkthroughs, a local Gradio demo, DeepSpeed finetune scripts, and reproducible evaluation code. Supported languages span Python, JavaScript, Go, Rust, TypeScript, and dozens more. Repository code is MIT licensed, and the model license allows commercial use.
Software developers and ML engineers use DeepSeek Coder for in-editor completion, multi-file infilling, coding chat assistants, and fine-tuning on custom instruction datasets.
Four model sizes from 1B to 33B, each in base and instruct variants
16K context window built for multi-file, repository-level completion
Fill-in-the-blank pretraining supports inserting code between existing blocks
Covers 80+ languages including Python, JavaScript, Go, Rust, and Solidity
Finetune scripts with DeepSpeed plus vLLM inference examples in the repo
Open model weights and MIT-licensed repo code with commercial use allowed.
Published benchmarks place it ahead of other open-source code models in several tests.
Multiple parameter sizes let teams trade inference speed against capability.
Larger 33B models need substantial GPU memory for local inference.
Documentation is centered on the GitHub README rather than a standalone docs site.
Instruct models require eos_token_id changes for optimal code completion behavior.
What is DeepSeek Coder?
DeepSeek Coder is a series of open-source code language models from DeepSeek. The models handle code completion, infilling, repository-level generation, and instruction-following coding tasks across many programming languages.
What model sizes does DeepSeek Coder offer?
DeepSeek Coder ships in 1B, 5.7B, 6.7B, and 33B parameter sizes. Each size has a base model for completion and an instruct model tuned for chat-style coding requests.
Is DeepSeek Coder free to use?
Yes. DeepSeek Coder model weights are available on Hugging Face, the repository code is MIT licensed, and the project states that commercial use is supported under the model license.
How do I run DeepSeek Coder locally?
DeepSeek Coder provides Python examples in its GitHub README using Hugging Face Transformers and PyTorch. You can also run a local Gradio demo from the demo folder or use vLLM for higher-throughput inference.
Which benchmarks does DeepSeek Coder perform well on?
DeepSeek Coder reports strong pass@1 scores on HumanEval, MultiPL-E, MBPP, DS-1000, and APPS. The README includes detailed evaluation tables and scripts in the Evaluation directory.
Can I fine-tune DeepSeek Coder on my own data?
Yes. DeepSeek Coder includes a finetune_deepseekcoder.py script with DeepSpeed support. Training data should follow the instruction and output JSONL format described in the finetune README.