{
  "quantization_method": "auto-round",
  "scheme": "W4A16",
  "bits": 4,
  "group_size": 128,
  "sym": true,
  "data_type": "int",
  "act_bits": 16,
  "iters": 1000,
  "nsamples": 500,
  "calibration_dataset": "NeelNanda/pile-10k (AutoRound default)",
  "calibration_type": "text-only (language model only)",
  "quantized_layers": 252,
  "fp16_layers": 105,
  "original_model": "TomoroAI/tomoro-colqwen3-embed-4b",
  "note": "Vision encoder kept in FP16 (not quantized). Text-only calibration is appropriate since only language_model is quantized."
}