Is DeepSeek-Coder open source?

Yes — deepseek-ai/DeepSeek-Coder is open source, released under the MIT license.

What language is DeepSeek-Coder written in?

deepseek-ai/DeepSeek-Coder is primarily written in Python.

How popular is DeepSeek-Coder?

deepseek-ai/DeepSeek-Coder has 23.9k stars on GitHub and is currently holding steady.

Where can I find DeepSeek-Coder?

deepseek-ai/DeepSeek-Coder is on GitHub at https://github.com/deepseek-ai/DeepSeek-Coder.

← all repositories

deepseek-ai/DeepSeek-Coder

A code model that reads your repo, not just your cursor

DeepSeek Coder exists to give open-source code models actual repository context, training on dependency-sorted files across a 16K window instead of isolated snippets.

★23.9k stars Python Coding Assistants Language Models

View on GitHub ↗ Homepage ↗

Velocity · 7d

+6.7

★ / day

Trend

→steady

star history

What it does

DeepSeek Coder is a family of open-source code language models ranging from 1B to 33B parameters, designed for code completion, insertion, and instruction-following. The models were trained from scratch on 2 trillion tokens—87% code and 13% natural language in English and Chinese—and support dozens of programming languages. Unlike simpler autocomplete tools, they handle project-level context using a 16K window and a fill-in-the-blank training objective.

The interesting bit

The training data isn’t just raw file dumps. The team parsed intra-repo file dependencies, rearranged files according to their dependency order, and concatenated them into single examples so the model learns cross-file relationships. They also used repo-level minhash deduplication rather than simple file-level deduplication, which suggests they actually care about the signal-to-noise ratio of repository context.

Key highlights

The 33B base model outperforms CodeLlama-34B by 7.9% on HumanEval Python, 9.3% on HumanEval Multilingual, 10.8% on MBPP, and 5.9% on DS-1000, according to the project’s benchmarks.
The 7B base model reaches parity with CodeLlama-34B, implying you can get 34B-like results on consumer hardware.
The 33B instruct variant outperforms GPT35-turbo on HumanEval and achieves comparable results on MBPP.
Supports repository-level completion by ingesting dependent files in dependency order, not just the current buffer.
Available in both base (completion/infilling) and instruct (chat/Q&A) variants.

Caveats

The README reports only pass@1 results, so multi-sample performance (pass@k) is unclear.
All benchmark claims are self-reported by the model authors; independent verification isn’t mentioned.
The repository is primarily a model release and demo wrapper, not a standalone IDE plugin or LSP server—you’ll need to wire it into your own tooling.

Verdict

Worth a look if you need an open-weight alternative to GitHub Copilot or CodeLlama and don’t mind hosting the inference yourself. Skip it if you were hoping for a plug-and-play IDE extension—this is the engine, not the car.

Frequently asked

What is deepseek-ai/DeepSeek-Coder?: DeepSeek Coder exists to give open-source code models actual repository context, training on dependency-sorted files across a 16K window instead of isolated snippets.
Is DeepSeek-Coder open source?: Yes — deepseek-ai/DeepSeek-Coder is open source, released under the MIT license.
What language is DeepSeek-Coder written in?: deepseek-ai/DeepSeek-Coder is primarily written in Python.
How popular is DeepSeek-Coder?: deepseek-ai/DeepSeek-Coder has 23.9k stars on GitHub and is currently holding steady.
Where can I find DeepSeek-Coder?: deepseek-ai/DeepSeek-Coder is on GitHub at https://github.com/deepseek-ai/DeepSeek-Coder.