Submission · qwen35-27b

`d791b318abab4f30`

by AxionML NVFP4 + FP8 KV (reference) · test split · runner kubernetes

NVFP4, ModelOpt + FP8 KV AxionML/Qwen3.5-27B-NVFP4 W4A4KV8

done

Pipeline

build done
contract done
quality done
efficiency done
score done

Result

1.777LowBScore

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

W·A·KV	W4A4KV8 declared · the service reads W4A4KV8
Accuracy	tier A · mean drop 1.10 points
Drift from BF16	KL 0.028 nats per token · generated length +0.8% · D = 1.12 points
Speed gain	1.90× BF16 in the same pod (serve 1.98×, interactive 1.60×, long context 2.16×) (gate: at least 0.95× less its error)
Memory gain	2.10× BF16's peak (serve 2.13×, interactive 2.26×, long context 1.91×) · 36.5 GiB against BF16's 77.8 GiB on the first workload
Efficiency gain E	1.99
Per-domain results on the hidden split are not released.

1.76 GPU-hours

Raw record

{
  "agent": "zoo-pub-axionml-nvfp4-kvfp8",
  "created_at": "2026-10-08T04:35:47Z",
  "error": null,
  "id": "d791b318abab4f30",
  "participant": "AxionML NVFP4 + FP8 KV",
  "partition": "reference",
  "result": {
    "admitted": true,
    "anchor_efficiency": {
      "serve": 842.233403691062
    },
    "bf16_peak_gib": 77.756,
    "drift": 1.1191,
    "efficiency": {
      "serve": 1669.6474495519165
    },
    "efficiency_gain": 1.9872,
    "gpu_hours": 1.756,
    "kl": {
      "kl": 0.027570758271263963,
      "kl_median": 0.024722817385142454,
      "kl_per_token": 0.027438725849248297,
      "kl_se": 0.0012491033506038325,
      "sequences": 256,
      "tokens": 63668
    },
    "length_shift": 0.00822,
    "mean_drop": 0.01103,
    "memories": {
      "interactive": 2.2578,
      "long_context": 1.913,
      "serve": 2.1326
    },
    "memory": 2.0962,
    "memory_saved": 0.5311,
    "metric": "lowbit_score",
    "peak_gib": 36.461,
    "peak_memory": {
      "serve": {
        "device_peak_gib": 161.72,
        "kv_bytes_per_token": 53145,
        "kv_gib": 13.786,
        "kv_usage": 0.0996,
        "other_gib": 4.785,
        "peak_gib": 36.461,
        "weights_gib": 17.89
      }
    },
    "precision": {
      "activations": 4,
      "kv": 8,
      "label": "W4A4KV8",
      "weights": 4,
      "weights_average": 4.9
    },
    "qualified": true,
    "quality": {
      "tier": "A",
      "valid_domains": 5
    },
    "reason": null,
    "score": 1.7768,
    "score_se": 0.0428,
    "speed": 1.8994,
    "speed_serve": 1.9824,
    "speeds": {
      "interactive": 1.5999,
      "long_context": 2.1604,
      "serve": 1.9824
    },
    "task_id": "qwen35-27b",
    "timed_audit": null
  },
  "runner": "kubernetes",
  "solution": {
    "checkpoint": "AxionML/Qwen3.5-27B-NVFP4",
    "declared_precision": true,
    "detected_precision": "W4A4KV8",
    "method": "NVFP4, ModelOpt + FP8 KV",
    "precision": "W4A4KV8"
  },
  "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": "2026-10-09T03:22:14Z"
}