---
license: gemma
datasets:
- agentlans/LinguaNova
base_model:
- google/gemma-3-4b-it
tags:
- gemma
- quantized
- llama.cpp
- multilingual
- instruction-tuned
- natural-language-processing
- calibration
- importance-matrix
- resource-efficient
language:
- multilingual
---
# Gemma 3 4B Instruct Quantized Models

This repo offers quantized versions of [google/gemma-3-4b-it](https://huggingface.co/google/gemma-3-4b-it) for use with llama.cpp.
Quantization was done using [an unofficial Docker image](https://hub.docker.com/r/agentlans/llamacpp) 
and calibrated on 100 rows from the [agentlans/LinguaNova](https://huggingface.co/datasets/agentlans/LinguaNova) dataset to maintain coherence and multilingual support. 
The importance matrix file is included.

## Limitations

- Optimized for multilingual natural language tasks.
- May underperform on math, coding, and untested multimodal features.
- Shares all limitations and biases of the original Gemma 3 models.

## Notes

- Ideal for resource-constrained environments.
- Test on your data for best results.
- See the original [google/gemma-3-4b-it](https://huggingface.co/google/gemma-3-4b-it) page for full details and guidelines.

*This card covers only the quantized models.*