Submission · qwen35-27b

`42398c749ac94d38`

by Qwen FP8 + FP8 KV (reference) · test split · runner kubernetes

FP8, 128x128 blocks + FP8 KV Qwen/Qwen3.5-27B-FP8 W8A8KV8

done

Pipeline

build done
contract done
quality done
efficiency done
score done

Result

pending LowBScore

speed or memory on the other workloads not measured

W·A·KV	W8A8KV8 declared · the service reads W8A8KV8
Accuracy	tier A · mean drop 0.21 points
Drift from BF16	KL 0.002 nats per token
Speed gain	1.04× BF16 in the same pod (serve 1.04×) (gate: at least 0.95× less its error)
Memory gain	— BF16's peak (serve 1.74×) · 44.7 GiB against BF16's 77.8 GiB on the first workload
Per-domain results on the hidden split are not released.

1.11 GPU-hours

Raw record

{
  "agent": "zoo-pub-qwen-fp8-kvfp8",
  "created_at": "2026-10-08T04:35:47Z",
  "error": null,
  "id": "42398c749ac94d38",
  "participant": "Qwen FP8 + FP8 KV",
  "partition": "reference",
  "result": {
    "admitted": true,
    "anchor_efficiency": {
      "serve": 859.1838800646054
    },
    "bf16_peak_gib": 77.756,
    "efficiency": {
      "serve": 894.9550849810769
    },
    "gpu_hours": 1.113,
    "kl": {
      "kl": 0.0023786130223120826,
      "kl_median": 0.0019354782102181732,
      "kl_per_token": 0.0023772559199434534,
      "kl_se": 0.0003019480989325232,
      "sequences": 256,
      "tokens": 63668
    },
    "mean_drop": 0.0021,
    "memories": {
      "interactive": null,
      "long_context": null,
      "serve": 1.7402
    },
    "memory": null,
    "memory_saved": 0.4253,
    "metric": "lowbit_score",
    "peak_gib": 44.683,
    "peak_memory": {
      "serve": {
        "device_peak_gib": 163.046,
        "kv_bytes_per_token": 53150,
        "kv_gib": 13.787,
        "kv_usage": 0.1058,
        "other_gib": 3.225,
        "peak_gib": 44.683,
        "weights_gib": 27.67
      }
    },
    "precision": {
      "activations": 8,
      "kv": 8,
      "label": "W8A8KV8",
      "weights": 8,
      "weights_average": 8.17
    },
    "qualified": false,
    "quality": {
      "tier": "A",
      "valid_domains": 5
    },
    "reason": "speed or memory on the other workloads not measured",
    "score": null,
    "score_se": null,
    "speed": null,
    "speed_serve": 1.0416,
    "speeds": {
      "interactive": null,
      "long_context": null,
      "serve": 1.0416
    },
    "task_id": "qwen35-27b",
    "timed_audit": null
  },
  "runner": "kubernetes",
  "solution": {
    "checkpoint": "Qwen/Qwen3.5-27B-FP8",
    "declared_precision": true,
    "detected_precision": "W8A8KV8",
    "method": "FP8, 128x128 blocks + FP8 KV",
    "precision": "W8A8KV8"
  },
  "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": "qwen35-27b",
  "track": "artifact",
  "updated_at": "2026-10-08T22:33:20Z"
}