{
  "id": "whisper-large-v3-turbo",
  "identity": {
    "family": "whisper",
    "display_name": "Whisper large-v3-turbo",
    "hub_id": "openai/whisper-large-v3-turbo",
    "api_name": null,
    "revision": null,
    "variants": [],
    "license": "MIT",
    "commercial_use": true,
    "commercial_use_note": "MIT weights and code; no use restrictions.",
    "release_date": "2024-10",
    "paper_url": "https://arxiv.org/abs/2212.04356",
    "repo_url": "https://github.com/openai/whisper"
  },
  "task": {
    "category": "transcription",
    "modalities_in": [
      "audio",
      "video"
    ],
    "modalities_out": [
      "text"
    ],
    "best_at": "Multilingual speech-to-text on GPU at large-v3 quality with the decoder cut from 32 to 4 layers, so decoding is much faster than large-v3.",
    "worst_at": "Word-level timing through the transformers pipeline (Live Nation measured 9.4 GB peak and wrong times); CPU-only clusters (the code forces the whisper-1 API when there is no GPU); music and non-speech audio."
  },
  "io": {
    "input": {
      "audio": {
        "sample_rate": 16000,
        "long_form": "return_timestamps=True (MVS_WHISPER_RETURN_TIMESTAMPS), needed past 30 s"
      },
      "language": "optional forced language; auto-detect when unset"
    },
    "output": {
      "text": "transcript per audio chunk"
    }
  },
  "mixpeek_wiring": {
    "extractors": [
      {
        "extractor": "multimodal_extractor@v1",
        "selected_by": "run_transcription=true and HARDWARE_WHISPER_USE_API=false on a GPU cluster"
      },
      {
        "extractor": "multimodal_extractor@v2",
        "selected_by": "run_transcription=true and HARDWARE_WHISPER_USE_API=false on a GPU cluster"
      }
    ],
    "feature_uris": [],
    "vector_indexes": [],
    "retriever_stages": [],
    "composes_with": [
      "video -> audio chunks -> whisper transcript -> E5 transcription vector (multimodal_extractor_v*_transcription_embedding)",
      "customers/livenation/manifest-loop ln_shots plugin: 30 s windows with segment timestamps, then Qwen3-Embedding-0.6B text vectors"
    ]
  },
  "run": {
    "status": "served",
    "status_reason": "Ray Serve app in the inference registry. multimodal_extractor loads it in-process (WhisperLocalTranscriptionBatch, GPU) only when the whisper-1 API is off; prod runs HARDWARE_WHISPER_USE_API=true per infra/hardware/config.py comments.",
    "serve_apps": [
      "openai__whisper_large_v3_turbo"
    ],
    "runtime": "transformers AutoModelForSpeechSeq2Seq + AutoProcessor + pipeline(\"automatic-speech-recognition\")",
    "weights_format": "safetensors",
    "trust_remote_code": false,
    "size_on_disk_mb": 1618,
    "params_millions": 809,
    "preprocessing": "Decode and resample audio to 16 kHz mono; fp16 on GPU, fp32 otherwise.",
    "postprocessing": "Keep result[\"text\"]; timestamps are computed and discarded.",
    "serving_shape": "both",
    "batch_size": null,
    "serve_batch": null,
    "scale_to_zero": {
      "min_replicas": 0,
      "max_replicas": 1,
      "idle_timeout_s": 600,
      "memory_gb": null,
      "num_cpus": 0.5,
      "num_gpus": {
        "expr": "float(os.getenv('HARDWARE_WHISPER_NUM_GPUS', '0.5')) if hardware_config.gpu_concurrency > 0 else 0"
      },
      "cold_start_s": {
        "warm_node_s": null,
        "new_node_s": null,
        "source": null
      },
      "cpu_fallback": false
    }
  },
  "cost": {
    "unmeasured": true,
    "measured": null,
    "note": "No committed full measurement. Code comment (BACKE-740): return_timestamps cuts L4 throughput from 2.54 to 1.30 (units not stated)."
  },
  "evidence": {
    "unmeasured": true,
    "items": []
  },
  "refs": [
    "hf://openai/whisper-large-v3-turbo@41f01f3fe87f28c78e2fbf8b568835947dd65ed9"
  ],
  "formats": [
    {
      "ref": "hf://openai/whisper-large-v3-turbo@41f01f3fe87f28c78e2fbf8b568835947dd65ed9",
      "format": "safetensors",
      "file": "model.safetensors",
      "loader": "transformers / open_clip from_pretrained; Mixpeek Hub import (api/models/huggingface.py)",
      "supported_today": true
    }
  ],
  "use_cases": [
    "transcripts"
  ],
  "recommended": {
    "value": true,
    "for": [
      "transcripts"
    ],
    "reason": "The served speech model behind multimodal_extractor transcription."
  },
  "demos": [],
  "seo": {
    "title": "Whisper large-v3-turbo: speech-to-text model",
    "description": "Whisper large-v3-turbo is a audio and video speech-to-text model (MIT). Served on Mixpeek, scaling to zero when idle.",
    "keywords": [
      "Whisper large-v3-turbo",
      "whisper",
      "speech-to-text",
      "openai/whisper-large-v3-turbo",
      "transcripts"
    ],
    "faqs": [
      {
        "q": "What is Whisper large-v3-turbo best at?",
        "a": "Multilingual speech-to-text on GPU at large-v3 quality with the decoder cut from 32 to 4 layers, so decoding is much faster than large-v3."
      },
      {
        "q": "What are the limits of Whisper large-v3-turbo?",
        "a": "Word-level timing through the transformers pipeline (Live Nation measured 9.4 GB peak and wrong times); CPU-only clusters (the code forces the whisper-1 API when there is no GPU); music and non-speech audio."
      },
      {
        "q": "Can Whisper large-v3-turbo be used commercially?",
        "a": "Yes. MIT weights and code; no use restrictions."
      },
      {
        "q": "How does Whisper large-v3-turbo run on Mixpeek?",
        "a": "Ray Serve app in the inference registry. multimodal_extractor loads it in-process (WhisperLocalTranscriptionBatch, GPU) only when the whisper-1 API is off; prod runs HARDWARE_WHISPER_USE_API=true per infra/hardware/config.py comments."
      }
    ]
  },
  "last_verified": "2026-11-05",
  "sources": [
    "https://huggingface.co/openai/whisper-large-v3-turbo",
    "https://huggingface.co/api/models/openai/whisper-large-v3-turbo",
    "https://arxiv.org/abs/2212.04356",
    "server/engine/inference/openai/whisper_large_v3_turbo/services.py",
    "server/engine/extractors/multimodal_extractor/v2/pipeline.py",
    "server/infra/hardware/config.py",
    "customers/livenation/manifest-loop/README.md",
    "customers/livenation/manifest-loop/plugin/ln_shots/v1/models.py"
  ],
  "registry": {
    "openai__whisper_large_v3_turbo": {
      "deployment": "openai__whisper_large_v3_turbo",
      "service_id": "openai/whisper-large-v3-turbo",
      "routes": "engine/inference/openai/whisper_large_v3_turbo/routes.py",
      "route_prefix": "/openai/whisper_large_v3_turbo/v1",
      "enabled": {
        "expr": "not _whisper_use_api"
      },
      "description": "OpenAI Whisper Large V3 Turbo speech recognition model. Transcribes audio to text with state-of-the-art accuracy. Supports multiple languages and audio formats.",
      "capabilities": {
        "category": "transcription",
        "modality_input": [
          "audio"
        ],
        "modality_output": "text",
        "supports_batch": false,
        "reranking_mode": null
      },
      "vector_config": null,
      "serve": {
        "min_replicas": 0,
        "max_replicas": 1,
        "idle_timeout_s": 600,
        "target_ongoing_requests": 5,
        "num_cpus": 0.5,
        "num_gpus": {
          "expr": "float(os.getenv('HARDWARE_WHISPER_NUM_GPUS', '0.5')) if hardware_config.gpu_concurrency > 0 else 0"
        },
        "memory_gb": null,
        "pinnable": false,
        "placement": "worker",
        "enabled": true
      },
      "interactive_twin": false,
      "interactive_serve": null,
      "schemas": {
        "input": "shared.inference.openai.whisper_large_v3_turbo.models.InferenceRequest",
        "output": "shared.inference.openai.whisper_large_v3_turbo.models.InferenceResponse",
        "config": "shared.inference.openai.whisper_large_v3_turbo.models.InferenceConfigs"
      }
    }
  },
  "templates": [],
  "usage": {
    "namespaces": null,
    "collections": null,
    "uncounted": true,
    "templates": 0,
    "counted_at": null,
    "source": null
  },
  "facets": {
    "status": "served",
    "category": "transcription",
    "family": "whisper",
    "license": "MIT",
    "commercial_use": true,
    "modalities_in": [
      "audio",
      "video"
    ],
    "modalities_out": [
      "text"
    ],
    "use_cases": [
      "transcripts"
    ],
    "recommended": true,
    "release_date": "2024-10",
    "params_millions": 809,
    "size_on_disk_mb": 1618,
    "vector_dims": null,
    "vector_distance": null,
    "weights_format": "safetensors",
    "formats": [
      "safetensors"
    ],
    "loadable_today": true,
    "trust_remote_code": false,
    "cpu_fallback": false,
    "num_gpus": null,
    "memory_gb": null,
    "min_replicas": 0,
    "serving_shape": "both",
    "served": false,
    "ref_schemes": [
      "hf"
    ],
    "has_evidence": false,
    "cost_measured": false,
    "template_count": 0,
    "usage_namespaces": null,
    "usage_collections": null,
    "usage_uncounted": true,
    "has_demo": false,
    "rank_score": 4
  },
  "rank_score": 4
}