Overview
DeepSeek Coder was the first model family shipped by Chinese AI lab DeepSeek, released on 2 November 2023. It is a series of open code language models trained from scratch on roughly 2 trillion tokens (87% source code and 13% natural language in English and Chinese), spanning more than 80 programming languages. The line comes in four sizes — DeepSeek-Coder-1.3B, 5.7B, 6.7B, and 33B — each shipped in a Base (pre-trained) and an Instruct (instruction-tuned) variant.
Every model uses a 16K-token context window and was pre-trained at the repository level with a fill-in-the-middle (FIM) objective, so it handles project-level code completion and infilling rather than just isolated snippets. At launch, DeepSeek-Coder-Base-33B was the strongest open code model of its size, and the instruction-tuned DeepSeek-Coder-Instruct-33B was reported to outperform GPT-3.5-turbo on HumanEval and match it on MBPP.
DeepSeek Coder has since been superseded: DeepSeek-Coder-V2 (a much larger Mixture-of-Experts model with a far longer context) launched in June 2024, and DeepSeek lists the original line as discontinued. The original weights remain downloadable on Hugging Face under the DeepSeek Model License, which permits commercial use, so the models are still useful as small, self-hostable code assistants.
| Released | 2023-11-02 |
|---|---|
| License | Code under MIT License; model weights under the DeepSeek Model License (permits commercial use with responsible-use restrictions). |
| Weights | Open weights |
| Parameters | Family: 1.3B, 5.7B (MoE), 6.7B, and 33B |
| Context | 16K tokens |
| Max output | Not separately specified — bounded by the 16K context window |
| Architecture | Decoder-only transformer code LLM, pre-trained from scratch on ~2T tokens (87% source code, 13% English/Chinese natural language) at the repository level with a 16K window and a fill-in-the-blank (FIM) objective; Instruct variants fine-tuned on instruction data. |
| Knowledge cutoff | Not officially published |
| Modalities | text, code |
| Status | Superseded — the original DeepSeek Coder line (Nov 2023) was replaced by DeepSeek-Coder-V2 in June 2024 and is listed by DeepSeek as discontinued. The open weights remain available for download and self-hosting. |
Benchmarks
- HumanEval (Python, Pass@1) — Instruct 33B79.3%
- HumanEval (Python, Pass@1) — Base 33B56.1%
- MBPP (Pass@1) — Instruct 33B70%
- MBPP (Pass@1) — Base 33B66%
- DS-1000 (Pass@1) — Instruct 33B40.2%
- HumanEval (Python, Pass@1) — Base 6.7B49.4%
- HumanEval (Python, Pass@1) — Base 1.3B34.8%
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Strengths
- Fully open weights with a license that allows commercial use, ideal for on-prem / air-gapped code assistants
- Small sizes (1.3B, 6.7B) run on a single consumer GPU, with the 33B for higher quality
- 16K context plus repository-level + fill-in-the-middle training suits code completion and infilling, not just chat
- Strong-for-its-era HumanEval/MBPP results; the 33B Instruct beat GPT-3.5-turbo on HumanEval at launch
- Broad language coverage — 80+ programming languages, including multilingual HumanEval evaluation
Best for
- Self-hosted code completion and infilling in an IDE or editor plugin
- Local / offline code generation where data cannot leave the network
- Fine-tuning a small, permissively-licensed base model for a domain-specific coding assistant
- Repository-level autocomplete and FIM tasks using the 16K window
- Research and benchmarking baselines for open code-LLM work
How to access
| Provider | Model ID |
|---|---|
| Hugging Face (deepseek-ai) ↗ | deepseek-ai/deepseek-coder-33b-instruct |
| Hugging Face (deepseek-ai) ↗ | deepseek-ai/deepseek-coder-6.7b-instruct |
| Ollama ↗ | deepseek-coder |
DeepSeek Coder — every version
The full lineage of the DeepSeek Coder line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| DeepSeek-Coder-V2current | 2024-06-17 | — | Open weights |
| DeepSeek Coder | 2023-11-02 | — | Open weights |
FAQ
Is DeepSeek Coder open source?
The weights are open and free to download on Hugging Face. The code repository is MIT-licensed, and the model weights are under the DeepSeek Model License, which permits commercial use subject to responsible-use restrictions. It is best described as open-weight (source-available) rather than a fully OSI-approved license.
What sizes does DeepSeek Coder come in?
Four sizes — 1.3B, 5.7B, 6.7B, and 33B parameters — each available as a Base (pre-trained) model and an Instruct (instruction-tuned) model. All share a 16K-token context window.
Is DeepSeek Coder still the latest version?
No. The original DeepSeek Coder was released in November 2023 and was superseded by DeepSeek-Coder-V2 in June 2024, which uses a Mixture-of-Experts architecture and a much longer context. DeepSeek lists the original line as discontinued, but the weights remain downloadable.
How well does DeepSeek Coder do on coding benchmarks?
Per the DeepSeek-Coder paper, the 33B Instruct model scores about 79.3% Pass@1 on HumanEval (Python) and 70% on MBPP — at launch it outperformed GPT-3.5-turbo on HumanEval. The 33B Base scores 56.1% on HumanEval, with smaller sizes scaling down (49.4% for 6.7B, 34.8% for 1.3B).