{
  "schema_version": 1,
  "family": "parakeet_tdt",
  "display_name": "Parakeet-TDT 0.6B v3",
  "description": "NVIDIA Parakeet-TDT 0.6B v3 FastConformer-TDT ASR covering 25 European languages with automatic language detection. Supports the upstream Transformers-compatible safetensors package and standalone audio.cpp GGUF, with offline full-context, bounded-window long-form, and buffered streaming; the checkpoint uses unlimited bidirectional attention and is not a native cache-aware streaming model.",
  "category": "asr",
  "status": "community",
  "tasks": [
    "asr"
  ],
  "modes": [
    "offline",
    "streaming"
  ],
  "languages": [
    "bg",
    "cs",
    "da",
    "de",
    "el",
    "en",
    "es",
    "et",
    "fi",
    "fr",
    "hr",
    "hu",
    "it",
    "lt",
    "lv",
    "mt",
    "nl",
    "pl",
    "pt",
    "ro",
    "ru",
    "sk",
    "sl",
    "sv",
    "uk"
  ],
  "runtime": {
    "tags": [
      "gguf",
      "stream"
    ]
  },
  "capabilities": {
    "asr": [
      "word_timestamps",
      "partial_results"
    ]
  },
  "options": {
    "request": [
      {
        "name": "max_tokens",
        "type": "int",
        "description": "Maximum TDT generated tokens; 0 or omitted uses the model-derived limit.",
        "required": false,
        "min": 0,
        "default": 0
      },
      {
        "name": "keep_language_tags",
        "type": "bool",
        "description": "Keep language tag tokens in decoded text; default false.",
        "required": false,
        "default": false
      }
    ],
    "session": [
      {
        "name": "weight_type",
        "type": "enum",
        "description": "Shared matmul weight storage type; default native.",
        "preset": "weight_type_full",
        "required": false,
        "default": "native"
      },
      {
        "name": "matmul_weight_type",
        "type": "enum",
        "description": "Encoder and decoder matmul weight storage type; defaults to weight_type, which defaults to native. Q8_0 measured 1.79x faster on the tested CPU and changed roughly 8 percent of transcripts without moving aggregate word error rate.",
        "preset": "weight_type_full",
        "required": false
      },
      {
        "name": "conv_weight_type",
        "type": "enum",
        "description": "Convolution weight storage type; default native.",
        "preset": "weight_type_conv",
        "required": false,
        "default": "native"
      },
      {
        "name": "perf_mode",
        "type": "enum",
        "description": "Encoder attention implementation. Default off uses the validated relative-attention path; flash_attention enables the fused implementation, which was numerically validated but slower on the tested hardware.",
        "preset": "perf_mode_flash_attention",
        "required": false,
        "default": "off"
      },
      {
        "name": "weight_context_mb",
        "type": "int",
        "description": "Weight context arena size in MiB; default 3072.",
        "required": false,
        "min": 1,
        "default": 3072
      },
      {
        "name": "encoder_graph_arena_mb",
        "type": "int",
        "description": "Encoder graph arena size in MiB; default 1024.",
        "required": false,
        "min": 1,
        "default": 1024
      },
      {
        "name": "decoder_graph_arena_mb",
        "type": "int",
        "description": "Decoder graph arena size in MiB; default 256.",
        "required": false,
        "min": 1,
        "default": 256
      },
      {
        "name": "audio_chunk_duration_sec",
        "type": "float",
        "description": "Center-region duration for buffered streaming in seconds; default 2. Fixed context windows are re-encoded rather than cache-aware.",
        "required": false,
        "min": 0.001,
        "default": 2.0
      },
      {
        "name": "left_context_sec",
        "type": "float",
        "description": "Past context included when re-encoding each buffered-streaming window in seconds; default 10.",
        "required": false,
        "min": 0.0,
        "default": 10.0
      },
      {
        "name": "right_context_sec",
        "type": "float",
        "description": "Future lookahead included when re-encoding each buffered-streaming window in seconds; default 2 and adds equivalent partial-result latency.",
        "required": false,
        "min": 0.0,
        "default": 2.0
      },
      {
        "name": "streaming_attention_mode",
        "type": "enum",
        "description": "Attention policy inside each buffered window. full_context preserves bidirectional attention over the bounded window.",
        "values": [
          "full_context"
        ],
        "required": false,
        "default": "full_context"
      },
      {
        "name": "offline_mode",
        "type": "enum",
        "description": "Offline encoder scheduling. full_context encodes the whole utterance, long_form uses bounded overlapping windows, and auto selects long_form beyond audio_chunk_threshold_sec.",
        "values": [
          "full_context",
          "long_form",
          "auto"
        ],
        "required": false,
        "default": "full_context"
      },
      {
        "name": "audio_chunk_threshold_sec",
        "type": "float",
        "description": "Duration threshold used by offline_mode=auto before switching to bounded-window long-form execution; default 30 seconds.",
        "required": false,
        "min": 0.001,
        "default": 30.0
      }
    ],
    "load": []
  },
  "package_defaults": {
    "download": {
      "kind": "huggingface_snapshot",
      "repo": "audio-cpp/audio.cpp-gguf",
      "revision": "main",
      "gated": false
    }
  },
  "packages": [
    {
      "id": "parakeet_tdt_q8_0",
      "display_name": "Parakeet-TDT 0.6B v3 Q8_0 GGUF",
      "default": true,
      "format": "gguf",
      "precision": "q8_0",
      "target_directory": "Parakeet-TDT-0.6B-v3-GGUF",
      "files": [
        "Parakeet-TDT-0.6B-v3-GGUF/parakeet-tdt-0.6b-v3-q8_0.gguf"
      ],
      "strip_prefix": "Parakeet-TDT-0.6B-v3-GGUF"
    },
    {
      "id": "parakeet_tdt_f16",
      "display_name": "Parakeet-TDT 0.6B v3 F16 GGUF",
      "format": "gguf",
      "precision": "f16",
      "target_directory": "Parakeet-TDT-0.6B-v3-GGUF",
      "files": [
        "Parakeet-TDT-0.6B-v3-GGUF/parakeet-tdt-0.6b-v3-f16.gguf"
      ],
      "strip_prefix": "Parakeet-TDT-0.6B-v3-GGUF"
    }
  ],
  "dependencies": [],
  "ui": {
    "recommended_package": "parakeet_tdt_q8_0",
    "tags": [
      "ASR",
      "GGUF"
    ],
    "docs": [
      "docs/community_models/parakeet_tdt.md"
    ]
  },
  "sources": [
    {
      "format": "gguf",
      "roots": {
        "model": ".",
        "weights": "$gguf"
      },
      "files": {
        "config": "model:config.json",
        "processor_config": "model:processor_config.json",
        "tokenizer_json": "model:tokenizer.json"
      },
      "tensors": {
        "weights": "weights:"
      }
    },
    {
      "format": "safetensors",
      "roots": {
        "model": "."
      },
      "files": {
        "config": "model:config.json",
        "processor_config": "model:processor_config.json",
        "tokenizer_json": "model:tokenizer.json"
      },
      "tensors": {
        "weights": "model:model.safetensors"
      }
    }
  ]
}