Llama-3.2-3B-Thai_CRM - GGUF
This repository contains GGUF quantized weights for yothinS/Llama-3.2-3B-Thai_CRM, a model fine-tuned for Customer Relationship Management (CRM) workflows, Thai business communications, and customer behavioral analysis.
GGUF files can be run locally on CPU, Apple Silicon, or NVIDIA/AMD GPUs via llama.cpp, Ollama, LM Studio, or carrany/text-generation-webui.
Available Files & Quantization Comparison
| Filename | Quant Type | File Size | Recommended RAM/VRAM | Quality / Use Case |
|---|---|---|---|---|
Llama-3.2-3B-Thai_CRM.Q2_K.gguf | 2-bit (Q2_K) | 1.36 GB | ~3.0 GB | Significant quality loss; ultra low-memory testing only. |
Llama-3.2-3B-Thai_CRM.Q3_K_M.gguf | 3-bit (Q3_K_M) | 1.69 GB | ~3.5 GB | Low resource usage; noticeable loss in complex reasoning. |
Llama-3.2-3B-Thai_CRM.Q4_K_M.gguf | 4-bit (Q4_K_M) | 2.02 GB | ~4.0 GB | Recommended balance of quality, generation speed, and size. |
Llama-3.2-3B-Thai_CRM.Q5_K_M.gguf | 5-bit (Q5_K_M) | 2.32 GB | ~4.5 GB | High fidelity; preserves nuance in Thai token generation. |
Llama-3.2-3B-Thai_CRM.Q6_K.gguf | 6-bit (Q6_K) | 2.64 GB | ~5.0 GB | Very close to original 16-bit performance with minor savings. |
Llama-3.2-3B-Thai_CRM.Q8_0.gguf | 8-bit (Q8_0) | 3.42 GB | ~5.8 GB | Near-identical output to unquantized FP16/BF16 weights. |
Recommendation: Start with
Q4_K_Mfor general deployment, orQ5_K_Mif you require higher accuracy on complex Thai sentence parsing.
How to Run
1. Using llama.cpp CLI
Download your chosen GGUF file using the Hugging Face CLI:
huggingface-cli download yothinS/Llama-3.2-3B-Thai_CRM-GGUF Llama-3.2-3B-Thai_CRM.Q4_K_M.gguf --local-dir .