We have quantized the meta-llama/Llama-3.2-3B-Instruct model into three variants:
Q5_KM
Q4_KM
IQ4_XS
These quantized models offer improved efficiency while maintaining performance.
Discover our full range of quantized language models by visiting our SandLogic Lexicon GitHub.
To learn more about our company and services, check out our website at SandLogic.
Model Type: Multilingual large language model (LLM)
Architecture: Auto-regressive language model with optimized transformer architecture
Parameters: 3 billion
Training Approach: Supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF)
Data Freshness: Pretraining data cutoff of December 2023
Model Capabilities
Llama-3.2-3B-Instruct is optimized for multilingual dialogue use cases, including:
Agentic retrieval
Summarization tasks
Assistant-like chat applications
Knowledge retrieval
Query and prompt rewriting
Intended Use
Commercial and research applications in multiple languages
Mobile AI-powered writing assistants
Natural language generation tasks (with further adaptation)
Training Data
Pretrained on up to 9 trillion tokens from publicly available sources
Incorporates knowledge distillation from larger Llama 3.1 models
Fine-tuned with human-generated and synthetic data for safety
Safety Considerations
Implements safety mitigations as in Llama 3
Emphasis on appropriate refusals and tone in responses
Includes safeguards against borderline and adversarial prompts
Quantized Variants
Q5_KM: 5-bit quantization using the KM method
Q4_KM: 4-bit quantization using the KM method
IQ4_XS: 4-bit quantization using the IQ4_XS method
These quantized models aim to reduce model size and improve inference speed while maintaining performance as close to the original model as possible.
Usage
pip install llama-cpp-python
Please refer to the llama-cpp-python documentation to install with GPU support.
Basic Text Completion
Here's an example demonstrating how to use the high-level API for basic text completion:
bash
1from llama_cpp import Llama
23llm = Llama(4model_path="./models/7B/Llama-3.2-3B-Instruct-Q5_K_M.gguf",
5verbose=False,
6# n_gpu_layers=-1, # Uncomment to use GPU acceleration7# n_ctx=2048, # Uncomment to increase the context window8)910output = llm.create_chat_completion(11 messages =[12{13"role":"system",
14"content":"You are a pirate chatbot who always responds in pirate speak!",
15},
16{"role":"user", "content":"Who are you?"},
17]18)1920print(output["choices"][0]['message']['content'])
Download
You can download Llama models in gguf format directly from Hugging Face using the from_pretrained method. This feature requires the huggingface-hub package.
By default, from_pretrained will download the model to the Hugging Face cache directory. You can manage installed model files using the huggingface-cli tool.
Acknowledgements
We thank Meta for developing the original Llama-3.2-3B-Instruct model.
Special thanks to Georgi Gerganov and the entire llama.cpp development team for their outstanding contributions.
