Submission · qwen3-8b

`eaeaddae97584b8d`

by Qwen/Qwen3-8B-FP8 (reference) · test split · runner kubernetes

FP8, 128x128 blocks Qwen/Qwen3-8B-FP8 W8A8KV16

done

Pipeline

build done
contract done
quality done
efficiency done
score done

Result

1.173LowBScore

Qualified: the accuracy, speed, and length gates passed.

W·A·KV	W8A8KV16
Accuracy	tier A · mean drop 0.15 points
Drift from BF16	KL 0.006 nats per token · generated length +2.2% · D = 0.17 points
Speed gain	1.19× BF16 in the same pod (serve 1.17×, interactive 1.19×, long context 1.20×) (gate: at least 0.95× less its error)
Memory gain	1.23× BF16's peak (serve 1.13×, interactive 1.57×, long context 1.04×) · 49.2 GiB against BF16's 55.5 GiB on the first workload
Efficiency gain E	1.19
Per-domain results on the hidden split are not released.

0.56 GPU-hours

Raw record

{
  "agent": "hf.Qwen.Qwen3-8B-FP8",
  "created_at": "2026-10-07T18:11:04Z",
  "error": null,
  "id": "eaeaddae97584b8d",
  "participant": "Qwen/Qwen3-8B-FP8",
  "partition": "reference",
  "result": {
    "admitted": true,
    "anchor_efficiency": {
      "serve": 1895.3811941842116
    },
    "bf16_peak_gib": 55.544,
    "drift": 0.1675,
    "efficiency": {
      "serve": 2211.078737287171
    },
    "efficiency_gain": 1.1928,
    "gpu_hours": 0.561,
    "kl": {
      "kl": 0.005677251780486823,
      "kl_median": 0.00463960230806135,
      "kl_per_token": 0.006033537653904628,
      "kl_se": 0.001072587354753993,
      "sequences": 64,
      "tokens": 15418
    },
    "length_shift": 0.02218,
    "mean_drop": 0.00147,
    "memories": {
      "interactive": 1.5691,
      "long_context": 1.0417,
      "serve": 1.1281
    },
    "memory": 1.2263,
    "memory_gain": 1.128,
    "memory_saved": 0.1135,
    "metric": "lowbit_score",
    "peak_gib": 49.237,
    "peak_memory": {
      "serve": {
        "device_peak_gib": 163.124,
        "kv_bytes_per_token": 147270,
        "kv_gib": 38.202,
        "kv_usage": 0.2545,
        "other_gib": 2.155,
        "peak_gib": 49.237,
        "weights_gib": 8.88
      }
    },
    "precision": {
      "activations": 8,
      "kv": 16,
      "label": "W8A8KV16",
      "weights": 8,
      "weights_average": 8.0
    },
    "qualified": true,
    "quality": {
      "tier": "A",
      "valid_domains": 5
    },
    "reason": null,
    "score": 1.173,
    "score_se": 0.0304,
    "speed": 1.186,
    "speed_serve": 1.1666,
    "speeds": {
      "interactive": 1.1925,
      "long_context": 1.1993,
      "serve": 1.1666
    },
    "task_id": "qwen3-8b",
    "timed_audit": null
  },
  "runner": "kubernetes",
  "solution": {
    "checkpoint": "Qwen/Qwen3-8B-FP8",
    "declared_precision": false,
    "detected_precision": "W8A8KV16",
    "method": "FP8, 128x128 blocks",
    "precision": "W8A8KV16"
  },
  "split": "test",
  "stages": [
    {
      "name": "build",
      "status": "done"
    },
    {
      "name": "contract",
      "status": "done"
    },
    {
      "name": "quality",
      "status": "done"
    },
    {
      "name": "efficiency",
      "status": "done"
    },
    {
      "name": "score",
      "status": "done"
    }
  ],
  "status": "done",
  "task_id": "qwen3-8b",
  "track": "artifact",
  "updated_at": "2026-10-09T03:22:14Z"
}