Question 1

What is DeepSeek-Coder?

Accepted Answer

DeepSeek-Coder is an open-source series of code-focused language models from DeepSeek-AI, trained from scratch on 2T tokens (87% code, 13% English/Chinese text) for advanced code generation and completion.

Question 2

Which programming languages does DeepSeek-Coder support?

Accepted Answer

It covers 86+ programming languages in the original release and up to 338 languages in DeepSeek-Coder-V2.

Question 3

What context window sizes are available?

Accepted Answer

DeepSeek-Coder supports a 16K context window for project-level code completion and infilling, extended to 128K in V2.

Question 4

How does DeepSeek-Coder perform on benchmarks?

Accepted Answer

It achieves state-of-the-art performance among open-source models on HumanEval, MultiPL-E, MBPP, DS-1000, and APPS.

Question 5

What model sizes are offered?

Accepted Answer

Model sizes include 1B, 5.7B, 6.7B, and 33B, plus a Mixture-of-Experts DeepSeek-Coder-V2 with performance comparable to GPT-4-Turbo on coding and math tasks.

Question 6

What is special about DeepSeek-Coder-V2?

Accepted Answer

V2 adopts a Mixture-of-Experts architecture, expands language coverage and context to 128K, and targets strong efficiency and accuracy in coding/math tasks.

Question 7

Does DeepSeek-Coder support infilling or fill-in-the-middle?

Accepted Answer

Yes. It supports infilling (fill-in-the-blank) for editing and completing code within existing files and repositories.

Question 8

Is DeepSeek-Coder open-source?

Accepted Answer

Yes. DeepSeek-AI releases models and infrastructure openly; check the DeepSeek-AI GitHub organization and model repositories for licenses and usage terms.

Question 9

Where can I learn more or get the models?

Accepted Answer

Visit the DeepSeek-AI GitHub organization (deepseek-ai) and the official website (deepseek.com), or contact service@deepseek.com for details.

Question 10

What infrastructure tools are related to DeepSeek-AI’s models?

Accepted Answer

DeepSeek-AI also open-sources infrastructure such as 3FS (a high-performance distributed file system) and DeepEP (an EP communication library) for training and inference at scale.

Question 11