quant_meta.json

{
  "model_id": "HuggingFaceTB/SmolVLM-256M-Instruct",
  "family": "smolvlm",
  "method": "gptq_idefics3_merge",
  "bits": 4,
  "group_size": 128,
  "quant_time_s": 66.8,
  "n_calib": 128,
  "notes": "Inner Llama text decoder GPTQ-quantized, then re-attached to full Idefics3ForConditionalGeneration. Vision encoder + connector remain FP16."
}