{
  "schema_version": 1,
  "family": "nemotron_transcribe",
  "display_name": "NVIDIA Nemotron 3.5 Transcribe 30B-A4B",
  "description": "Experimental community port of the Nemotron-H speech language model for offline transcription, translation, audio questions, language identification, word timestamps and speaker labels.",
  "category": "asr",
  "status": "experimental",
  "tasks": ["asr"],
  "modes": ["offline"],
  "languages": ["auto"],
  "capabilities": {"asr": ["word_timestamps"]},
  "runtime": {"tags": ["gguf"]},
  "options": {
    "request": [
      {"name": "instruct", "type": "string", "required": false, "description": "Audio instruction or question. Replaces the template instruction; defaults to text_input when present. Chat controls and audio locators are rejected."},
      {"name": "template_name", "type": "enum", "values": ["default", "translation", "audio_qa", "language_id"], "default": "default", "required": false, "description": "Instruction template. Audio Q&A requires instruct or text_input. Timestamps and speakers require the default transcription template."},
      {"name": "language", "type": "string", "default": "auto", "required": false, "description": "Source language hint for the default transcription prompt. An explicit instruct replaces this prompt, including its language hint."},
      {"name": "target_language", "type": "string", "default": "English", "required": false, "description": "Target language for the translation template."},
      {"name": "system_prompt", "type": "string", "required": false, "description": "Optional system message; the trained default is empty."},
      {"name": "return_timestamps", "type": "bool", "default": false, "required": false, "description": "Run the checkpoint CTC head for source-language word timestamps. Failed streams return no fabricated word rows."},
      {"name": "diarization", "type": "bool", "default": false, "required": false, "description": "Request speaker-tagged transcription, CTC words and speaker activity turns. An explicit instruct must request speaker tags itself."},
      {"name": "max_tokens", "type": "int", "min": 1, "required": false, "description": "Maximum generated response tokens. Defaults to generation_config or native duration budget max(24,ceil(6*audio_seconds)+24). Prompt and output must fit max_context."},
      {"name": "do_sample", "type": "bool", "default": false, "required": false, "description": "Sample tokens instead of greedy decoding."},
      {"name": "temperature", "type": "float", "default": 1, "required": false, "description": "Positive sampling temperature when do_sample is true."},
      {"name": "top_k", "type": "int", "min": 0, "default": 0, "required": false, "description": "Sampling candidate count; zero disables top-k filtering."},
      {"name": "top_p", "type": "float", "min": 0, "max": 1, "default": 1, "required": false, "description": "Nucleus probability when sampling."},
      {"name": "repetition_penalty", "type": "float", "default": 1, "required": false, "description": "Positive token repetition penalty."},
      {"name": "seed", "type": "int", "min": 0, "default": 0, "required": false, "description": "Sampling seed."}
    ],
    "session": [
      {"name": "disk_experts", "type": "bool", "default": false, "required": false, "description": "Keep routed expert weights in the GGUF/host memory and stage selected experts through a bounded backend cache. Dense weights and all expert computation remain on the selected backend."},
      {"name": "pin_experts", "type": "bool", "default": false, "required": false, "description": "With disk_experts, retain complete native expert banks in pinned host RAM for direct uploads. Requires pinned host-buffer support."},
      {"name": "expert_cache_size", "type": "int", "min": 1, "default": 8, "required": false, "description": "Resident expert slots per routed layer with disk_experts. Must cover the model's experts per token and cannot exceed its expert count."},
      {"name": "disk_expert_read_threads", "type": "int", "min": 1, "max": 64, "default": 4, "required": false, "description": "Host workers for staging GGUF expert ranges when disk_experts is true and pin_experts is false. Parallel reads require OpenMP or Apple dispatch."},
      {"name": "weight_type", "type": "enum", "preset": "weight_type_full", "default": "native", "required": false, "description": "Native preserves mixed packaged tensor types. Changing this may greatly increase model memory."},
      {"name": "max_context", "type": "int", "min": 1, "default": 73728, "required": false, "description": "Combined text, audio and output context, bounded by backbone positions. Matches native SALM serving default."},
      {"name": "prefill_batch_size", "type": "int", "min": 1, "max": 2048, "default": 64, "required": false, "description": "Text/audio rows per decoder prefill graph."},
      {"name": "flash_attention", "type": "bool", "default": true, "required": false, "description": "Use flash attention in the speech and speaker encoders."},
      {"name": "encoder_window_size", "type": "int", "min": 1, "max": 5000, "default": 500, "required": false, "description": "Core encoder window in 80 ms output frames; the native default is 40 seconds."},
      {"name": "encoder_left_context", "type": "int", "min": 0, "max": 5000, "default": 50, "required": false, "description": "Encoder window left context in 80 ms frames."},
      {"name": "encoder_right_context", "type": "int", "min": 0, "max": 5000, "default": 50, "required": false, "description": "Encoder window right context in 80 ms frames."},
      {"name": "speaker_cache_size", "type": "int", "min": 16, "max": 5000, "default": 188, "required": false, "description": "Native serving speaker cache in 80 ms frames, distinct from training configuration. At least two rows per speaker are required."},
      {"name": "speaker_fifo_size", "type": "int", "min": 0, "max": 5000, "default": 40, "required": false, "description": "Recent speaker FIFO size in 80 ms frames."},
      {"name": "speaker_update_period", "type": "int", "min": 1, "max": 5000, "default": 300, "required": false, "description": "Native serving speaker-cache update period in 80 ms frames."}
    ],
    "load": []
  },
  "package_defaults": {
    "download": {
      "kind": "huggingface_snapshot",
      "repo": "Jmica/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-GGUF",
      "revision": "304032f14e1e9e4042f5491d2d0194c044710bb7",
      "gated": false
    }
  },
  "packages": [
    {
      "id": "nemotron_transcribe_iq4_nl",
      "display_name": "Nemotron 3.5 Transcribe 30B-A4B IQ4_NL experts",
      "default": true,
      "format": "gguf",
      "precision": "iq4_nl",
      "target_directory": "NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-GGUF",
      "files": [
        "nemotron-transcribe-30b-a4b-experts-IQ4_NL.gguf",
        "mmproj-NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-Encoder-F16.gguf"
      ]
    },
    {
      "id": "nemotron_transcribe_q4_k",
      "display_name": "Nemotron 3.5 Transcribe 30B-A4B Q4_K experts",
      "format": "gguf",
      "precision": "q4_k",
      "target_directory": "NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-GGUF",
      "files": [
        "nemotron-transcribe-30b-a4b-experts-Q4_K_S.gguf",
        "mmproj-NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-Encoder-F16.gguf"
      ]
    },
    {
      "id": "nemotron_transcribe_q3_k",
      "display_name": "Nemotron 3.5 Transcribe 30B-A4B Q3_K experts",
      "format": "gguf",
      "precision": "q3_k",
      "target_directory": "NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-GGUF",
      "files": [
        "nemotron-transcribe-30b-a4b-experts-Q3_K_S.gguf",
        "mmproj-NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-Encoder-F16.gguf"
      ]
    },
    {
      "id": "nemotron_transcribe_q2_k",
      "display_name": "Nemotron 3.5 Transcribe 30B-A4B Q2_K experts",
      "format": "gguf",
      "precision": "q2_k",
      "target_directory": "NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-GGUF",
      "files": [
        "nemotron-transcribe-30b-a4b-experts-Q2_K.gguf",
        "mmproj-NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-Encoder-F16.gguf"
      ]
    }
  ],
  "dependencies": [],
  "ui": {"recommended_package": "nemotron_transcribe_iq4_nl", "tags": ["GGUF"], "docs": ["docs/models/nemotron_transcribe.md"]},
  "sources": [
    {
      "format": "gguf",
      "roots": {"model": ".", "weights": "$gguf"},
      "files": {
        "backbone_config": "model:llm_backbone/config.json",
        "generation_config": "model:generation_config.json",
        "processor_config": "model:processor_config.json",
        "frontend_window": "model:frontend_window.json",
        "tokenizer_json": "model:tokenizer.json",
        "tokenizer_config": "model:tokenizer_config.json"
      },
      "tensors": {"weights": {"source": "weights:"}}
    }
  ]
}