{
 "schema_version": "nanojev-mlx-quant-v1",
 "source_checkpoint": "NanoJev",
 "source_weights_sha256": "fff62d1412685c1714eaa386acb603f9690371fb3cc8ad03dc41319302597c28",
 "bits": 4,
 "group_size": 64,
 "include_embed": true,
 "scales_dtype": "float16",
 "note": "backbone linears + embed_tokens in MLX affine 4-bit (group 64); head and norms kept in fp16",
 "quantized_params_by_bucket": {
  "mlp": 264241152,
  "self_attn": 176160768,
  "embed_tokens": 155582464
 },
 "relative_max_error_by_bucket": {
  "embed_tokens": {
   "max": 0.06698610633611679,
   "mean": 0.06698610633611679
  },
  "mlp": {
   "max": 0.12449125945568085,
   "mean": 0.06786140923698743
  },
  "self_attn": {
   "max": 0.11749649047851562,
   "mean": 0.07736335305630096
  }
 },
 "packed_weights_bytes": 335857708,
 "packed_weights_sha256": "332d47eb23c81b0370f580e87e0e55ef737ba1baa206a8d300eadf9617307b62",
 "dequant_fp32_bytes": 2384507132,
 "dequant_fp32_sha256": "2e5bf94be4fa2264e95e4ee3ba763d6240a41faaff6f6cd643f9e1eea29fe291",
 "seconds": 87.46
}