---
license: gemma
base_model: google/translategemma-27b-it
base_model_relation: quantized
pipeline_tag: translation
library_name: transformers
tags:
- translation
- subtitles
- compressed-tensors
- nvfp4
- vllm
---

# henry1477/translategemma-27b-it-NVFP4A16

This is a public, text-only `NVFP4A16` derivative produced with a GPTQ recipe from
[`google/translategemma-27b-it`](https://huggingface.co/google/translategemma-27b-it) at immutable revision
`7d10f0b72f89a2d0f268cea30727d8b77c0d25c2`. The vision tower and multimodal projector were removed; the
published architecture is `Gemma3ForCausalLM` and `lm_head` was excluded from
quantization because Gemma 3 ties it to the embedding table. The upstream multimodal checkpoint has 27.43B parameters; this text-only derivative has 27.01B.

## Quantization and calibration

- `llm-compressor` GPTQ with `NVFP4A16` compressed-tensors weights.
- 512 production-shaped JSON-scene prompts, rendered through TranslateGemma's
  own chat template, with source languages `ar bg ca cs da de el es fi fr he hi hu id it ja ko ms nl no pl pt ro ru sv th tr uk vi zh` and English/Latin
  American Spanish targets. During quantization, llm-compressor reported a numerical-Hessian fallback to round-to-nearest for 62 module(s); the exact count is disclosed here and the release gates test the emitted weights.
- The upstream processor/chat template and generation configuration are
  preserved, including EOS token IDs `[1, 106]`.
- The release benchmark used max context 4096, constrained
  JSON output, GPU memory utilization 0.78, max
  sequences 4, and wave concurrency
  2.

## Intended use

This checkpoint is intended for vLLM subtitle translation on NVIDIA Blackwell.
Use constrained JSON decoding and independently validate cue IDs. The release
evidence below is the measured 12B-versus-27B real-corpus comparison for this
exact base revision; it is not copied from the earlier 12B checkpoint.

```python
from vllm import LLM

llm = LLM(
    model="henry1477/translategemma-27b-it-NVFP4A16",
    max_model_len=4096,
    gpu_memory_utilization=0.78,
    max_num_seqs=4,
    limit_mm_per_prompt={"image": 0},
)
```

## Release evidence

```json
{
  "approval": {
    "passed": true,
    "reason": "All structural, two-wave release, multilingual, source-conditioned QE, and adversarial manual-audit gates passed."
  },
  "artifacts": {
    "manual-quality-audit.json": "d00f8c6afbbff1d5d1dcc904ddf9248057fdbd6df23ebe4600d0e5cef3c5b59d",
    "metricx-quality.json": "865adfd33054d57f8fb1ac40f3d016d66c5c46d65b31aaad0985c503411254b6",
    "multi-12b.json": "393a466316c496f138ca6a3000e34ac49f1770becc40ba5a5d7f728f67f868aa",
    "multi-27b.json": "d6802aae0ec472ad4db06d95248e60e04c0426f7d634ebcd3f6de2f01e8ffe58",
    "production-waves.json": "31f28efa6a4d3424f21241026dd3f32935c67fa41c4d3417b4f7a402da1dc8cf",
    "real-12b.json": "69e6e69d8594f5250a3b3e199ed45798fca35f94b8a373da351b69413c29709f",
    "real-27b-wave1.json": "0d148ac727dfb3fd51c61599d235dcd2a10718a26e166fe43a4a8813059c9db6",
    "real-27b-wave2.json": "334e59c75deba46a47ce3da607bad7cb639925fb208ad660bfe3523569b61d75"
  },
  "baseline": {
    "metrics": {
      "chrf_samples": 261,
      "chrf_vs_int8": 95.86,
      "chrf_vs_reference": 95.86,
      "clock_drift_pct": 1395.9,
      "constrained_json": true,
      "copy_rate": 1.9157088122605364,
      "cues": 261,
      "empty_cues": 0,
      "engine": "vllm",
      "first_token_seconds": null,
      "generate_seconds": 20.5,
      "generation_requests": 18,
      "label": "12b-real-corpus-production-recovery",
      "length_ratio_median": 1.091,
      "load_seconds": 97.9,
      "max_model_len": 4096,
      "max_num_seqs": 16,
      "model": "/models/huggingface/models--henry1477--translategemma-12b-it-NVFP4A16/snapshots/a4afd9372aaad181f97ba2ba057524ea95d281e7",
      "output_tokens": 10146,
      "output_tokens_per_second": 494.1,
      "peak_vram_mib": 30456,
      "recovery_splits": 2,
      "scenes": 14,
      "scenes_per_minute": 40.91,
      "timestamp": "2026-08-28T04:51:20Z",
      "validated": 14,
      "validation_rate": 100.0
    },
    "repo": "henry1477/translategemma-12b-it-NVFP4A16",
    "revision": "a4afd9372aaad181f97ba2ba057524ea95d281e7",
    "structuralPass": true,
    "succeeded": true
  },
  "calibration": {
    "cues": 10240,
    "maxSequenceLength": 2048,
    "prompts": 512,
    "rtnFallbackModules": 62,
    "rtnFallbackPattern": "model.layers.{0..61}.mlp.down_proj",
    "sha256": "d5e42cebbf1e5c9f05e04c84751bac3076ca8886c98cad42442115075625159c",
    "sourceLanguages": [
      "ar",
      "bg",
      "ca",
      "cs",
      "da",
      "de",
      "el",
      "es",
      "fi",
      "fr",
      "he",
      "hi",
      "hu",
      "id",
      "it",
      "ja",
      "ko",
      "ms",
      "nl",
      "no",
      "pl",
      "pt",
      "ro",
      "ru",
      "sv",
      "th",
      "tr",
      "uk",
      "vi",
      "zh"
    ],
    "targetLanguages": [
      "en",
      "es-MX"
    ]
  },
  "candidate": {
    "manualReview": {
      "baselineMajorErrorCueIds": [
        "a1-c000031",
        "a1-c000129",
        "a1-c000137"
      ],
      "candidateMajorErrorCueIds": [
        "a1-c000071",
        "a1-c000193"
      ],
      "counts": {
        "baseline_better": 9,
        "candidate_better": 10,
        "equivalent_or_no_material_regression": 31
      },
      "limitations": "Source ASR contains many incomplete fragments; judgments assess faithfulness at the durable cue boundary and explicitly retain ambiguous cases as equivalent rather than guessing missing context. This is an AI-assisted manual audit, not a human review.",
      "method": "Adversarial sample: the 20 largest MetricX candidate-minus-baseline error deltas plus systematic every-eighth-cue coverage to 50 unique cues. Each Spanish ASR fragment was compared directly with both English outputs.",
      "passed": true,
      "reviewedAt": "2026-08-28T05:08:00Z",
      "reviewerType": "Codex bilingual source-grounded manual audit",
      "sampledCues": 50
    },
    "modelLoaded": true,
    "multilingual": {
      "chrfVs12BBySource": {
        "ar": 72.42,
        "bg": 76.13,
        "ca": 77.91,
        "cs": 74.33,
        "da": 86.96,
        "de": 70.92,
        "el": 81.63,
        "es": 88.2,
        "fi": 85.81,
        "fr": 81.35,
        "he": 75.67,
        "hi": 70.25,
        "hu": 84.41,
        "id": 77.8,
        "it": 83.87,
        "ja": 58.54,
        "ko": 59.12,
        "ms": 60.5,
        "nl": 83.55,
        "no": 81.41,
        "pl": 79.38,
        "pt": 75.32,
        "ro": 82.33,
        "ru": 80.34,
        "sv": 79.61,
        "th": 59.11,
        "tr": 71.0,
        "uk": 69.63,
        "vi": 76.03,
        "zh": 63.02
      },
      "chrfVs12BMean": 75.55,
      "chrfVs12BMedian": 83.23,
      "cues": 1200,
      "emptyCues": 0,
      "generationSeconds": 143.34,
      "label": "27b-multilingual",
      "loadSeconds": 131.22,
      "model": "/out/translategemma-27b-it-NVFP4A16",
      "peakVramMiB": 26625,
      "productionProfile": {
        "constrainedJson": true,
        "gpuMemoryUtilization": 0.78,
        "maxModelLen": 4096,
        "maxNumSeqs": 4
      },
      "scenes": 60,
      "schemaVersion": 1,
      "sourceCopyRate": 2.92,
      "sourceLanguages": [
        "ar",
        "bg",
        "ca",
        "cs",
        "da",
        "de",
        "el",
        "es",
        "fi",
        "fr",
        "he",
        "hi",
        "hu",
        "id",
        "it",
        "ja",
        "ko",
        "ms",
        "nl",
        "no",
        "pl",
        "pt",
        "ro",
        "ru",
        "sv",
        "th",
        "tr",
        "uk",
        "vi",
        "zh"
      ],
      "structuralPassRate": 100.0,
      "structuralScenes": 60,
      "targets": [
        "en",
        "es-MX"
      ],
      "timestamp": "2026-08-28T05:17:11Z"
    },
    "qualityRegression": false,
    "realCorpusWaves": [
      {
        "chrf_samples": 261,
        "chrf_vs_int8": 65.19,
        "chrf_vs_reference": 65.19,
        "clock_drift_pct": 180.5,
        "constrained_json": true,
        "copy_rate": 2.681992337164751,
        "cues": 261,
        "empty_cues": 0,
        "engine": "vllm",
        "first_token_seconds": null,
        "generate_seconds": 38.7,
        "generation_requests": 14,
        "label": "27b-real-corpus-wave-1",
        "length_ratio_median": 1.061,
        "load_seconds": 187.3,
        "max_model_len": 4096,
        "max_num_seqs": 4,
        "model": "/out/translategemma-27b-it-NVFP4A16",
        "output_tokens": 8292,
        "output_tokens_per_second": 214.52,
        "peak_vram_mib": 24800,
        "recovery_splits": 0,
        "scenes": 14,
        "scenes_per_minute": 21.73,
        "timestamp": "2026-08-28T04:55:26Z",
        "validated": 14,
        "validation_rate": 100.0
      },
      {
        "chrf_samples": 261,
        "chrf_vs_int8": 64.86,
        "chrf_vs_reference": 64.86,
        "clock_drift_pct": 1392.3,
        "constrained_json": true,
        "copy_rate": 2.681992337164751,
        "cues": 261,
        "empty_cues": 0,
        "engine": "vllm",
        "first_token_seconds": null,
        "generate_seconds": 37.7,
        "generation_requests": 14,
        "label": "27b-real-corpus-wave-2",
        "length_ratio_median": 1.062,
        "load_seconds": 135.0,
        "max_model_len": 4096,
        "max_num_seqs": 4,
        "model": "/out/translategemma-27b-it-NVFP4A16",
        "output_tokens": 8276,
        "output_tokens_per_second": 219.65,
        "peak_vram_mib": 26618,
        "recovery_splits": 0,
        "scenes": 14,
        "scenes_per_minute": 22.29,
        "timestamp": "2026-08-28T05:00:06Z",
        "validated": 14,
        "validation_rate": 100.0
      }
    ],
    "releaseStable": true,
    "sourceConditionedQuality": {
      "baselineMeanError": 5.573672558992773,
      "baselineWins": 104,
      "candidateMeanError": 4.607441142943627,
      "candidateWins": 140,
      "cues": 261,
      "evaluator": {
        "codeRevision": "fc4978eb064670f7cc33e93ea4f52d38396b8ae6",
        "repo": "google/metricx-24-hybrid-large-v2p6",
        "revision": "51e875ba5c525c81627cfd135ee10f43c87dce00"
      },
      "inputSha256": "728959634edee1956351111a951b0c9112a37c43780ae76db9b73ebbace9153a",
      "lowerIsBetter": true,
      "mode": "reference-free-QE",
      "outputSha256": "6654017948d897bd135ab9841a10c18c3c36c0b40bb1c3315b572242c1ff01a2",
      "ties": 17
    },
    "structuralPass": true,
    "waves": [
      {
        "elapsedSeconds": 195.98844933509827,
        "executed": true,
        "exitCode": 0,
        "generationCalls": 8,
        "number": 1,
        "releaseStable": true,
        "remainingGpuProcesses": "",
        "structuralPass": true
      },
      {
        "elapsedSeconds": 53.89741849899292,
        "executed": true,
        "exitCode": 0,
        "generationCalls": 8,
        "number": 2,
        "releaseStable": true,
        "remainingGpuProcesses": "",
        "structuralPass": true
      }
    ]
  },
  "environment": {
    "cudaRuntime": "13.2",
    "driverVersion": "595.79",
    "gpuName": "NVIDIA GeForce RTX 5090",
    "gpuUuid": "GPU-4bd4db57-095d-d266-8453-a4d9dc822610",
    "pipelineCommit": "7cde44cede87fa97cf5e57125c002ea33af3aee9",
    "torchVersion": "2.13.0+cu130",
    "vllmVersion": "0.27.1",
    "vramMiB": 32607
  },
  "model": {
    "architecture": "Gemma3ForCausalLM",
    "baseRepo": "google/translategemma-27b-it",
    "baseRevision": "7d10f0b72f89a2d0f268cea30727d8b77c0d25c2",
    "checkpointSha256": "a636a4d1b71c6e0c7d8450e0d70c2ed494fa74594b4a8652c0da2f24b7f92f5c",
    "parametersAfter": 27009346304,
    "parametersBefore": 27432406640,
    "repo": "henry1477/translategemma-27b-it-NVFP4A16",
    "scheme": "NVFP4A16"
  },
  "multilingualBaseline": {
    "chrfVs12BBySource": {},
    "chrfVs12BMean": null,
    "chrfVs12BMedian": null,
    "cues": 1200,
    "emptyCues": 0,
    "generationSeconds": 36.6,
    "label": "12b-multilingual",
    "loadSeconds": 114.06,
    "model": "/models/huggingface/models--henry1477--translategemma-12b-it-NVFP4A16/snapshots/a4afd9372aaad181f97ba2ba057524ea95d281e7",
    "peakVramMiB": 30993,
    "productionProfile": {
      "constrainedJson": true,
      "gpuMemoryUtilization": 0.88,
      "maxModelLen": 4096,
      "maxNumSeqs": 16
    },
    "scenes": 60,
    "schemaVersion": 1,
    "sourceCopyRate": 2.92,
    "sourceLanguages": [
      "ar",
      "bg",
      "ca",
      "cs",
      "da",
      "de",
      "el",
      "es",
      "fi",
      "fr",
      "he",
      "hi",
      "hu",
      "id",
      "it",
      "ja",
      "ko",
      "ms",
      "nl",
      "no",
      "pl",
      "pt",
      "ro",
      "ru",
      "sv",
      "th",
      "tr",
      "uk",
      "vi",
      "zh"
    ],
    "structuralPassRate": 100.0,
    "structuralScenes": 60,
    "targets": [
      "en",
      "es-MX"
    ],
    "timestamp": "2026-08-28T05:12:18Z"
  },
  "productionProfile": {
    "constrainedJson": true,
    "gpuMemoryUtilization": 0.78,
    "kvAllocatorHeadroom": true,
    "maxModelLen": 4096,
    "maxNumSeqs": 4,
    "waveConcurrency": 2
  },
  "qualityDecision": {
    "baselineStructuralPass": true,
    "candidateStructuralPass": true,
    "chrfVs12B": 65.19,
    "copyRatePass": true,
    "humanReviewPass": true,
    "lexicalAgreementPass": false,
    "manualReviewPass": true,
    "matchingRevisions": true,
    "minimumChrfVs12B": 80.0,
    "noQualityRegression": true,
    "noStructuralRegression": true,
    "present": true,
    "qualityMetricPass": true,
    "sourceConditionedQualityPass": true
  },
  "schemaVersion": 1
}
```

## License and provenance

This derivative is governed by the Gemma Terms of Use. You must review and
accept the gated upstream terms on
[`google/translategemma-27b-it`](https://huggingface.co/google/translategemma-27b-it); downloading this public
quantization does not waive or replace those terms. Provenance is pinned to
`google/translategemma-27b-it@7d10f0b72f89a2d0f268cea30727d8b77c0d25c2`.