DeepSeek Coder 33B Instruct AWQ
| Property | Value |
|---|---|
| Parameter Count | 33B |
| Quantization | 4-bit AWQ |
| Context Length | 16,384 tokens |
| License | DeepSeek License |
| Model Size | 18.01 GB |
What is deepseek-coder-33B-instruct-AWQ?
DeepSeek Coder 33B Instruct AWQ is a highly optimized coding assistant model that combines the power of the original DeepSeek 33B architecture with efficient 4-bit AWQ quantization. Trained on 2T tokens with an 87% code and 13% natural language split, this model specializes in computer science and programming tasks while maintaining a strict boundary against non-technical queries.
Implementation Details
The model utilizes Advanced Weight Quantization (AWQ) technology to compress the original 33B parameter model into a more efficient 4-bit format, resulting in an 18.01 GB file size. It maintains the original 16K token context window, allowing for comprehensive project-level code analysis and generation.
- Optimized with 128-group size AWQ quantization
- Trained on Evol Instruct Code dataset
- Supports multiple inference frameworks including vLLM, Text Generation Inference, and AutoAWQ
Core Capabilities
- Project-level code completion and infilling
- Multi-language programming support
- State-of-the-art performance on coding benchmarks
- Specialized technical Q&A focused on computer science
Frequently Asked Questions
Q: What makes this model unique?
This model combines DeepSeek's advanced coding capabilities with efficient 4-bit quantization, making it both powerful and resource-efficient while maintaining high performance on coding tasks. The strict focus on computer science-related queries ensures high-quality technical responses.
Q: What are the recommended use cases?
The model excels at code completion, technical documentation, algorithm implementation, and debugging assistance. It's particularly suitable for professional developers and computing environments with GPU acceleration support.
