DeepSeek Coder is a suite of open-source code language models trained from scratch on 2 trillion tokens, comprising 87% code and 13% natural language. It offers models ranging from 1B to 33B parameters, designed for project-level code completion and infilling with a 16K context window. The models achieve state-of-the-art performance on various coding benchmarks, making them suitable for developers and researchers seeking advanced code generation capabilities.

How It Works

The models are pre-trained on a massive dataset of code and natural language, with a focus on project-level context. This is achieved through a 16K context window and a fill-in-the-blank task, enabling the models to understand and generate code across entire projects. Instruction-tuned variants are also available for conversational coding assistance.

Quick Start & Requirements

• Install dependencies: pip install -r requirements.txt
• Requires PyTorch and Hugging Face Transformers.
• Demo available on Hugging Face Spaces.
• Official models can be downloaded from Hugging Face Hub.

Highlighted Details

• Outperforms existing open-source code LLMs on HumanEval, MBPP, and DS-1000 benchmarks.
• 7B model achieves performance comparable to CodeLlama-34B.
• 33B instruct model rivals GPT-3.5-turbo on HumanEval.
• Supports over 100 programming languages.

Maintenance & Community

• Active development and community support.
• Links to Discord and WeChat for community interaction.
• Resources available via awesome-deepseek-coder.

Licensing & Compatibility

• Code repository licensed under MIT.
• Model usage subject to a separate Model License.
• Supports commercial use.

Limitations & Caveats

• Quantization to GGUF (llama.cpp) and GPTQ (exllamav2) requires specific setup steps and potential PR merges.
• Instruct models can perform code completion with a specific eos_token_id adjustment.