quantization_config.json

{
  "base_model": "google/gemma-3-4b-it",
  "quantization": "int4_per_group_symmetric",
  "group_size": 128,
  "bits": 4,
  "method": "static_int4_dequantized",
  "description": "INT4 per-group symmetric quantization of LLM backbone weights. Vision encoder kept in fp16. Weights stored as dequantized fp16 for maximum compatibility.",
  "quantized_layers": 239,
  "skipped_vision_layers": 162
}