{
  "family": "qwen3_asr",
  "display_name": "Qwen3-ASR",
  "description": "Qwen ASR model family for language identification and speech recognition across 30 languages, 22 Chinese dialects, and multiple English accents, with robustness for noisy, long-form, and singing audio.",
  "category": "asr",
  "status": "supported",
  "tasks": [
    "asr"
  ],
  "modes": [
    "offline",
    "streaming"
  ],
  "languages": [
    "zh",
    "en",
    "yue",
    "ar",
    "de",
    "fr",
    "es",
    "pt",
    "id",
    "it",
    "ko",
    "ru",
    "th",
    "vi",
    "ja",
    "tr",
    "hi",
    "ms",
    "nl",
    "sv",
    "da",
    "fi",
    "pl",
    "cs",
    "fil",
    "fa",
    "el",
    "hu",
    "mk",
    "ro",
    "zh dialects"
  ],
  "capabilities": {
    "asr": [
      "word_timestamps",
      "vad_chunking",
      "partial_results"
    ]
  },
  "options": {
    "request": [
      {
        "name": "language",
        "type": "string",
        "description": "Recognition language code or label. Omit to detect language automatically; Auto also selects automatic detection.",
        "required": false
      },
      {
        "name": "max_tokens",
        "type": "int",
        "description": "Maximum generated transcript tokens; omitted uses the checkpoint generation_config.json (512 when absent).",
        "required": false,
        "min": 1
      },
      {
        "name": "return_timestamps",
        "type": "bool",
        "description": "Return word-level timestamps by running Qwen3 ForcedAligner after ASR; default false and requires forced_aligner_path when enabled. Offline only; streaming rejects true even when an aligner is configured.",
        "required": false,
        "default": false
      },
      {
        "name": "audio_chunk_mode",
        "type": "enum",
        "description": "Audio chunking mode; default auto uses fixed chunks normally and VAD chunks when timestamps are requested. Offline only; streaming uses fixed-duration windows regardless of this selection.",
        "values": [
          "auto",
          "fixed",
          "vad",
          "none"
        ],
        "required": false,
        "default": "auto"
      },
      {
        "name": "audio_chunk_duration_sec",
        "type": "float",
        "description": "Maximum audio chunk duration in seconds; must be > 0 when chunking or streaming. Ignored with audio_chunk_mode=none offline. Default 30 for streaming and offline transcription, or 15 for offline timestamps/VAD chunking.",
        "required": false,
        "min": 0
      },
      {
        "name": "clamp_timestamps_to_audio",
        "type": "bool",
        "description": "Opt-in guard for word timestamp output: keep repaired forced-aligner word spans inside the local audio chunk. Defaults to false to preserve existing timestamp repair behavior.",
        "required": false,
        "default": false
      },
      {
        "name": "preserve_punctuation",
        "type": "bool",
        "description": "Opt-in text output mode for timestamped chunked ASR: preserve ASR punctuation in text_output instead of rebuilding text from aligned words.",
        "required": false,
        "default": false
      }
    ],
    "session": [
      {
        "name": "perf_mode",
        "type": "enum",
        "values": [
          "off",
          "cuda_perf"
        ],
        "description": "Performance preset. cuda_perf selects flash attention, packed projections, f16 KV, 256-token attention alignment and two independent offline chunks per decode batch. Explicit session options override the preset.",
        "required": false,
        "default": "off"
      },
      {
        "name": "attention",
        "type": "enum",
        "values": [
          "mixed",
          "auto",
          "eager",
          "flash"
        ],
        "description": "Decoder attention preference. Default mixed preserves manual prefill and flash cached decoding; cuda_perf selects flash for both. auto uses the shared backend policy. Explicit values override the preset.",
        "required": false
      },
      {
        "name": "packed_projections",
        "type": "bool",
        "description": "Pack decoder QKV and gate/up projections without changing weight dtype. Default false, true with cuda_perf. Explicit values override the preset.",
        "required": false
      },
      {
        "name": "kv_cache_type",
        "type": "enum",
        "values": [
          "f32",
          "f16"
        ],
        "description": "Decoder KV storage. Default f32, f16 with cuda_perf. Explicit values override the preset.",
        "required": false
      },
      {
        "name": "kv_cache_alignment",
        "type": "int",
        "min": 1,
        "description": "Cached decoder attention prefix alignment in tokens. Default 1 attends over full cache capacity; cuda_perf uses cached 256-token prefix windows. Explicit values override the preset.",
        "required": false
      },
      {
        "name": "offline_batch_size",
        "type": "int",
        "min": 1,
        "max": 2,
        "description": "Independent offline chunks decoded together. Default 1, 2 with cuda_perf. Does not change chunk boundaries or streaming scheduling. Explicit values override the preset.",
        "required": false
      },
      {
        "name": "lora",
        "type": "path",
        "description": "PEFT LoRA adapter directory or checkpoint file; merged into matching encoder/decoder weights at session creation.",
        "required": false
      },
      {
        "name": "lora_scale",
        "type": "float",
        "description": "Multiplier for the adapter's alpha/r scaling; default 1.",
        "required": false,
        "default": 1.0
      },
      {
        "name": "forced_aligner_path",
        "type": "path",
        "description": "Optional Qwen3 ForcedAligner model path required for return_timestamps.",
        "required": false
      },
      {
        "name": "vad_model_path",
        "type": "path",
        "description": "Optional VAD model path for VAD chunking (any built-in VAD family); default assets/framework/models/silero_vad.",
        "required": false,
        "default": "assets/framework/models/silero_vad"
      },
      {
        "name": "weight_type",
        "type": "enum",
        "description": "Shared text decoder weight storage fallback; default native.",
        "required": false,
        "default": "native",
        "preset": "weight_type_full"
      },
      {
        "name": "audio_encoder_weight_type",
        "type": "enum",
        "description": "Audio encoder weight storage type; default native.",
        "required": false,
        "default": "native",
        "preset": "weight_type_conv"
      },
      {
        "name": "text_decoder_weight_type",
        "type": "enum",
        "description": "Text decoder matmul weight storage type; defaults to weight_type when set, otherwise native.",
        "required": false,
        "preset": "weight_type_full"
      },
      {
        "name": "audio_encoder_graph_arena_mb",
        "type": "int",
        "description": "Audio encoder graph arena size in MiB; default 128.",
        "required": false,
        "min": 1,
        "default": 128
      },
      {
        "name": "text_decoder_prefill_graph_arena_mb",
        "type": "int",
        "description": "Text decoder prefill graph arena size in MiB; default 256.",
        "required": false,
        "min": 1,
        "default": 256
      },
      {
        "name": "text_decoder_decode_graph_arena_mb",
        "type": "int",
        "description": "Text decoder cached-step graph arena size in MiB; default 256.",
        "required": false,
        "min": 1,
        "default": 256
      },
      {
        "name": "text_decoder_weight_context_mb",
        "type": "int",
        "description": "Text decoder weight context size in MiB; default 64.",
        "required": false,
        "min": 1,
        "default": 64
      }
    ],
    "load": []
  },
  "runtime": {
    "tags": [
      "gguf",
      "stream"
    ]
  },
  "ui": {
    "recommended_package": "qwen3_asr_1_7b_q8_0",
    "tags": [
      "ASR",
      "GGUF",
      "Stream"
    ],
    "docs": [
      "docs/models/qwen3_asr.md",
      "docs/asr.md",
      "docs/gguf.md"
    ]
  },
  "package_defaults": {
    "download": {
      "kind": "huggingface_snapshot",
      "repo": "audio-cpp/audio.cpp-gguf",
      "revision": "main",
      "gated": false
    }
  },
  "packages": [
    {
      "id": "qwen3_asr_1_7b_q8_0",
      "display_name": "Qwen3-ASR 1.7B Q8_0 GGUF",
      "default": true,
      "format": "gguf",
      "precision": "q8_0",
      "target_directory": "Qwen3-ASR-1.7B-GGUF",
      "files": [
        "Qwen3-ASR-1.7B-GGUF/qwen3-asr-1.7b-q8_0.gguf"
      ],
      "strip_prefix": "Qwen3-ASR-1.7B-GGUF"
    },
    {
      "id": "qwen3_asr_1_7b_f16",
      "display_name": "Qwen3-ASR 1.7B F16 GGUF",
      "format": "gguf",
      "precision": "f16",
      "target_directory": "Qwen3-ASR-1.7B-GGUF",
      "files": [
        "Qwen3-ASR-1.7B-GGUF/qwen3-asr-1.7b-f16.gguf"
      ],
      "strip_prefix": "Qwen3-ASR-1.7B-GGUF"
    },
    {
      "id": "qwen3_asr_0_6b_q8_0",
      "display_name": "Qwen3-ASR 0.6B Q8_0 GGUF",
      "format": "gguf",
      "precision": "q8_0",
      "target_directory": "Qwen3-ASR-0.6B-GGUF",
      "files": [
        "Qwen3-ASR-0.6B-GGUF/qwen3-asr-0.6b-q8_0.gguf"
      ],
      "strip_prefix": "Qwen3-ASR-0.6B-GGUF"
    },
    {
      "id": "qwen3_asr_0_6b_f16",
      "display_name": "Qwen3-ASR 0.6B F16 GGUF",
      "format": "gguf",
      "precision": "f16",
      "target_directory": "Qwen3-ASR-0.6B-GGUF",
      "files": [
        "Qwen3-ASR-0.6B-GGUF/qwen3-asr-0.6b-f16.gguf"
      ],
      "strip_prefix": "Qwen3-ASR-0.6B-GGUF"
    },
    {
      "id": "qwen3_asr_1_7b_safetensors",
      "display_name": "Qwen3-ASR 1.7B HF Safetensors",
      "format": "safetensors",
      "precision": "native",
      "target_directory": "Qwen3-ASR-1.7B-hf",
      "files": [
        "config.json",
        "generation_config.json",
        "model.safetensors",
        "processor_config.json",
        "tokenizer_config.json",
        "tokenizer.json"
      ],
      "download": {
        "kind": "huggingface_snapshot",
        "repo": "Qwen/Qwen3-ASR-1.7B-hf"
      }
    },
    {
      "id": "qwen3_asr_0_6b_safetensors",
      "display_name": "Qwen3-ASR 0.6B Safetensors",
      "format": "safetensors",
      "precision": "native",
      "target_directory": "Qwen3-ASR-0.6B",
      "files": [
        "config.json",
        "generation_config.json",
        "model.safetensors",
        "preprocessor_config.json",
        "tokenizer_config.json"
      ],
      "download": {
        "kind": "huggingface_snapshot",
        "repo": "Qwen/Qwen3-ASR-0.6B"
      }
    },
    {
      "id": "qwencleo_asr_q8_0",
      "display_name": "QwenCleo-ASR Egyptian Arabic Q8_0 GGUF",
      "description": "Egyptian Arabic and Arabic-English code-switching fine-tune of Qwen3-ASR-1.7B from mohammedaly22/QwenCleo-ASR. Use language=Arabic.",
      "format": "gguf",
      "precision": "q8_0",
      "target_directory": "QwenCleo-ASR-GGUF",
      "files": [
        "qwencleo-asr-q8_0.gguf"
      ],
      "download": {
        "kind": "huggingface_snapshot",
        "repo": "mohammedaly22/QwenCleo-ASR-GGUF"
      }
    },
    {
      "id": "qwencleo_asr_f16",
      "display_name": "QwenCleo-ASR Egyptian Arabic F16 GGUF",
      "description": "Egyptian Arabic and Arabic-English code-switching fine-tune of Qwen3-ASR-1.7B from mohammedaly22/QwenCleo-ASR. Use language=Arabic.",
      "format": "gguf",
      "precision": "f16",
      "target_directory": "QwenCleo-ASR-GGUF",
      "files": [
        "qwencleo-asr-f16.gguf"
      ],
      "download": {
        "kind": "huggingface_snapshot",
        "repo": "mohammedaly22/QwenCleo-ASR-GGUF"
      }
    },
    {
      "id": "qwencleo_asr_safetensors",
      "display_name": "QwenCleo-ASR Egyptian Arabic Safetensors",
      "description": "Egyptian Arabic and Arabic-English code-switching fine-tune of Qwen3-ASR-1.7B from mohammedaly22/QwenCleo-ASR. Use language=Arabic.",
      "format": "safetensors",
      "precision": "native",
      "target_directory": "QwenCleo-ASR",
      "files": [
        "config.json",
        "generation_config.json",
        "model.safetensors",
        "preprocessor_config.json",
        "tokenizer_config.json",
        "tokenizer.json",
        "vocab.json",
        "merges.txt",
        "chat_template.json"
      ],
      "download": {
        "kind": "huggingface_snapshot",
        "repo": "mohammedaly22/QwenCleo-ASR"
      }
    }
  ],
  "sources": [
    {
      "format": "gguf",
      "roots": {
        "model": ".",
        "weights": "$gguf"
      },
      "files": {
        "config": "model:config.json",
        "generation_config": "model:generation_config.json",
        "tokenizer_config": "model:tokenizer_config.json"
      },
      "optional_files": {
        "preprocessor_config": "model:preprocessor_config.json",
        "processor_config": "model:processor_config.json",
        "chat_template": "model:chat_template.json",
        "chat_template_jinja": "model:chat_template.jinja",
        "vocab": "model:vocab.json",
        "merges": "model:merges.txt",
        "tokenizer_json": "model:tokenizer.json"
      },
      "tensors": {
        "weights": "weights:"
      }
    },
    {
      "format": "safetensors",
      "roots": {
        "model": "."
      },
      "files": {
        "config": "model:config.json",
        "generation_config": "model:generation_config.json",
        "tokenizer_config": "model:tokenizer_config.json"
      },
      "optional_files": {
        "preprocessor_config": "model:preprocessor_config.json",
        "processor_config": "model:processor_config.json",
        "chat_template": "model:chat_template.json",
        "chat_template_jinja": "model:chat_template.jinja",
        "vocab": "model:vocab.json",
        "merges": "model:merges.txt",
        "tokenizer_json": "model:tokenizer.json"
      },
      "tensors": {
        "weights": "model:model.safetensors"
      }
    }
  ]
}