WeDLM-8B-Instruct-GGUF

First GGUF quantization of Tencent WeDLM-8B-Instruct!

Quantized using llama.cpp b7688.

Original model: tencent/WeDLM-8B-Instruct

About

WeDLM is an 8B parameter instruction-tuned model by Tencent, supporting English and Chinese. It features QK Norm architecture similar to Qwen3.

This GGUF uses qwen3 architecture identifier for maximum llama.cpp compatibility.

Available Files

Filename	Quant	Size	Description
WeDLM-8B-Instruct-Q4_K_M.gguf	Q4_K_M	4.68 GB	Good quality, recommended for most use cases
WeDLM-8B-Instruct-Q8_0.gguf	Q8_0	8.11 GB	High quality, best accuracy

Performance Benchmarks

CPU (16 threads, Zen4)

Quant	Prompt Processing	Text Generation
Q4_K_M	88.65 t/s	8.27 t/s
Q8_0	50.80 t/s	5.17 t/s

GPU (RTX 4060 Laptop, 8GB VRAM)

Quant	Prompt Processing	Text Generation
Q4_K_M	1833.84 t/s	37.08 t/s

Q4_K_M recommended for RTX 4060 (fits in 8GB VRAM)

Prompt Format (ChatML)

<|im_start|>system
You are a helpful AI assistant.<|im_end|>
<|im_start|>user
Hello!<|im_end|>
<|im_start|>assistant

Usage

llama.cpp

./llama-cli -m WeDLM-8B-Instruct-Q4_K_M.gguf \
  -p "<|im_start|>user\nHello<|im_end|>\n<|im_start|>assistant\n" \
  -n 256 -ngl 99

Ollama

# Create Modelfile
cat > Modelfile << 'EOF'
FROM ./WeDLM-8B-Instruct-Q4_K_M.gguf
TEMPLATE "<|im_start|>user\n{{ .Prompt }}<|im_end|>\n<|im_start|>assistant\n"
EOF

ollama create wedlm -f Modelfile
ollama run wedlm

Hardware Requirements

Quant	Min VRAM	Recommended RAM
Q4_K_M	6 GB	8 GB
Q8_0	10 GB	12 GB

Model Architecture

• Parameters: 8.19B
• Layers: 36
• Hidden Size: 4096
• Attention Heads: 32 (8 KV heads, GQA)
• Context Length: 16384
• Features: QK Norm, SwiGLU, RoPE (theta=1M)

Acknowledgements

• Original model: Tencent WeDLM Team
• Inference framework: llama.cpp

Disclaimer

This is an unofficial quantization. For official support, please refer to the original model repository.