{
  "source": "Contrastive-LM/CLM-v0.1-8B encoder = Qwen/Qwen3-8B",
  "source_revision": "b968826d9c46dd6066d109eabc6255188de91218",
  "tool": "mlx-lm 0.31.3 (mlx_lm.convert)",
  "command": "mlx_lm.convert --hf-path Qwen/Qwen3-8B --mlx-path <dir> -q --q-bits N --q-group-size 32",
  "mode": "affine, per-group bf16 scale and bias",
  "group_size": 32,
  "group_size_note": "32 is the published recipe and measured best at every width; mlx-lm's default (64) and Qwen's mlx-community repos (128) are both worse here. MLX affine spends 32 bits per group on a bf16 scale AND bias, so group size is a first-class quality knob.",
  "quantised": "every linear weight, including embed_tokens and lm_head; all of Qwen3-8B's quantized last dims (4096, 1024, 12288) divide 32, so nothing is skipped",
  "lm_head": "present in the safetensors for layout compatibility but replaced with an Identity before evaluation by clm_mlx.Encoder -- pooling never touches it",
  "sizes_bytes": {
    "4bit": 5131300029,
    "6bit": 7178920204,
    "8bit": 9226527200,
    "bf16": "unquantised reference, ~15 GB"
  },
  "sizes_note": "8190427136 x (bits + 32/group_size) / 8 + 616448 at group_size 32, i.e. bits + 1.0. Measured, not predicted: mlx_lm.convert reports 9.0117 / 7.0117 / 5.0118 bits per weight.",
  "variants": {
    "bf16": {
      "level": "reference",
      "recommended": true,
      "reason": "this is the bf16 baseline every other row is measured against; it is not itself a quantisation result",
      "top1": null,
      "decisive_top1": null,
      "gold_delta": null
    },
    "8bit": {
      "level": "recommended",
      "recommended": true,
      "reason": "meets the gate",
      "top1": 0.988757,
      "decisive_top1": 1.0,
      "decisive_n": 7440,
      "gold_delta": -0.004569,
      "gold_loss": 0.004569,
      "gold_tolerance": 0.005,
      "floor": 1.0
    },
    "6bit": {
      "level": "recommended",
      "recommended": true,
      "reason": "meets the gate",
      "top1": 0.98278,
      "decisive_top1": 1.0,
      "decisive_n": 7440,
      "gold_delta": 0.006958,
      "gold_loss": -0.006958,
      "gold_tolerance": 0.005,
      "floor": 1.0
    },
    "4bit": {
      "level": "not_recommended",
      "recommended": false,
      "reason": "agreement on decisive (>1 nat) decisions 0.889651 < 0.995; planner accuracy falls 31.95 points, beyond the 3.0-point budget",
      "top1": 0.618616,
      "decisive_top1": 0.889651,
      "decisive_n": 7440,
      "gold_delta": -0.319515,
      "gold_loss": 0.319515,
      "floor": 1.0
    }
  }
}