This repository provides GGUF quantized versions of the deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct model, optimized for local execution using llama.cpp and compatible ecosystems.

📌 Version Notes

All quantizations were generated from the official FP16 weights.

Target: Efficient execution on consumer hardware, mobile/edge devices, and systems with limited memory.
Performance: The output quality (reasoning, coherence, and accuracy) is strictly dependent on the base model's parameter scale (9B).

📊 Quantization Table

File	Method	Bit	Description
fp16.gguf	FP16	16-bit	Original Weights. No quantization applied. Maximum fidelity.
Q8_0.gguf	Q8_0	8-bit	Near-lossless. Practically identical to the original model with lower memory footprint.
Q5_K_M.gguf	Q5_K_M	5-bit	High Precision. Minimizes quantization error for critical tasks.
Q4_K_M.gguf	Q4_K_M	4-bit	Recommended. Best balance between speed and performance.
Q4_K_S.gguf	Q4_K_S	4-bit	Fast/Small. Optimized for maximum throughput and low RAM usage.

🛠️ Technical Details

Quantization Date: 2026-03-13
Tool used: llama-quantize (llama.cpp)
Method: K-Quantization (optimized for AVX2/AVX-512 and modern GPU architectures).

🚀 How to Use

Start a local OpenAI-compatible server with a web UI:

llama.cpp (CLI) using model from HuggingFace

./llama-cli -hf daniloreddy/DeepSeek-Coder-V2-Lite-Instruct_GGUF:Q4_K_M -p "User: Hello! Assistant:" -n 512 --temp 0.7

llama.cpp (CLI) using downloaded model

./llama-cli -m path/to/DeepSeek-Coder-V2-Lite-Instruct_Q4_K_M.gguf -p "User: Hello! Assistant:" -n 512 --temp 0.7

llama.cpp (SERVER) using model from HuggingFace

./llama-server -hf daniloreddy/DeepSeek-Coder-V2-Lite-Instruct_GGUF:Q4_K_M --port 8080 -c 4096

llama.cpp (SERVER) using downloaded model

./llama-server -m /path/to/DeepSeek-Coder-V2-Lite-Instruct_Q4_K_M.gguf --port 8080 -c 4096