DeepSeek Coder 6.7B Base AWQ
| Property | Value |
|---|---|
| Parameter Count | 6.7B |
| Quantization | 4-bit AWQ |
| Context Length | 16K tokens |
| License | DeepSeek License |
| Training Data | 2T tokens (87% code, 13% language) |
What is deepseek-coder-6.7B-base-AWQ?
DeepSeek Coder 6.7B Base AWQ is a quantized version of the original DeepSeek Coder model, specifically optimized for efficient code generation and completion tasks. This AWQ (Activation-aware Weight Quantization) variant reduces the model size while maintaining performance, making it more accessible for deployment on resource-constrained systems.
Implementation Details
The model has been quantized to 4-bit precision using AWQ technology, resulting in a compact 3.89GB size. It supports a substantial 16K token context window and is compatible with major inference frameworks including text-generation-webui, vLLM, and Hugging Face's TGI.
- Trained on 2 trillion tokens of diverse programming data
- Uses 128-group size quantization for optimal performance
- Supports project-level code completion and infilling
- Implements fill-in-the-blank task capabilities
Core Capabilities
- Code completion and generation across multiple programming languages
- Project-level code understanding and completion
- Code infilling and context-aware suggestions
- Support for both English and Chinese language interactions
- Efficient inference with 4-bit quantization
Frequently Asked Questions
Q: What makes this model unique?
This model combines the power of the original DeepSeek Coder with efficient 4-bit quantization, making it particularly suitable for production deployments while maintaining high-quality code generation capabilities.
Q: What are the recommended use cases?
The model excels in code completion, project-level development assistance, code infilling, and general programming tasks across multiple languages. It's particularly useful in scenarios where efficient resource usage is crucial while maintaining high-quality code generation capabilities.
