voice_audio_multi/model_specs/voxtral_realtime.json

{
  "family": "voxtral_realtime",
  "display_name": "Voxtral Mini 4B Realtime",
  "description": "Mistral 13-language realtime ASR model with a natively streaming causal audio encoder, configurable low-latency transcription delay, and accuracy competitive with offline open-source systems.",
  "category": "asr",
  "status": "supported",
  "tasks": [
    "asr"
  ],
  "modes": [
    "offline",
    "streaming"
  ],
  "languages": [
    "en",
    "zh",
    "hi",
    "es",
    "ar",
    "fr",
    "pt",
    "ru",
    "de",
    "ja",
    "ko",
    "it",
    "nl"
  ],
  "options": {
    "request": [
      {
        "name": "max_tokens",
        "type": "int",
        "min": 1,
        "description": "Maximum generated transcript tokens; omitted uses the audio-derived limit. Offline decoding only; ignored by live streaming.",
        "required": false
      },
      {
        "name": "do_sample",
        "type": "bool",
        "default": false,
        "description": "Sample rather than greedily decode.",
        "required": false
      },
      {
        "name": "temperature",
        "type": "float",
        "min": 0,
        "default": 1,
        "description": "Sampling temperature; must be > 0.",
        "required": false
      },
      {
        "name": "top_p",
        "type": "float",
        "min": 0,
        "max": 1,
        "default": 1,
        "description": "Nucleus sampling probability in (0, 1].",
        "required": false
      },
      {
        "name": "top_k",
        "type": "int",
        "min": 0,
        "default": 50,
        "description": "Top-k sampling limit; zero disables filtering.",
        "required": false
      },
      {
        "name": "seed",
        "type": "int",
        "min": 0,
        "default": 1234,
        "description": "Sampling seed.",
        "required": false
      }
    ],
    "session": [
      {
        "name": "weight_type",
        "type": "enum",
        "values": [
          "native",
          "f32",
          "f16",
          "bf16",
          "q4_0",
          "q4_1",
          "q5_0",
          "q5_1",
          "q2_k",
          "q3_k",
          "q4_k",
          "q5_k",
          "q6_k",
          "q8_0",
          "nvfp4"
        ],
        "default": "native",
        "description": "Shared weight storage type.",
        "required": false
      },
      {
        "name": "audio_encoder_weight_type",
        "type": "enum",
        "values": [
          "native",
          "f32",
          "f16",
          "bf16",
          "q4_0",
          "q4_1",
          "q5_0",
          "q5_1",
          "q2_k",
          "q3_k",
          "q4_k",
          "q5_k",
          "q6_k",
          "q8_0",
          "nvfp4"
        ],
        "description": "Audio encoder weight storage; inherits weight_type.",
        "required": false
      },
      {
        "name": "text_decoder_weight_type",
        "type": "enum",
        "values": [
          "native",
          "f32",
          "f16",
          "bf16",
          "q4_0",
          "q4_1",
          "q5_0",
          "q5_1",
          "q2_k",
          "q3_k",
          "q4_k",
          "q5_k",
          "q6_k",
          "q8_0",
          "nvfp4"
        ],
        "description": "Text decoder weight storage; inherits weight_type.",
        "required": false
      },
      {
        "name": "audio_encoder_graph_arena_mb",
        "type": "int",
        "min": 1,
        "default": 512,
        "description": "Audio encoder graph arena in MiB.",
        "required": false
      },
      {
        "name": "audio_encoder_weight_context_mb",
        "type": "int",
        "min": 1,
        "default": 128,
        "description": "Audio encoder weight context in MiB.",
        "required": false
      },
      {
        "name": "text_decoder_prefill_graph_arena_mb",
        "type": "int",
        "min": 1,
        "default": 512,
        "description": "Decoder prefill graph arena in MiB. Decoder uses the maximum of prefill and decode arena settings.",
        "required": false
      },
      {
        "name": "text_decoder_decode_graph_arena_mb",
        "type": "int",
        "min": 1,
        "default": 512,
        "description": "Decoder step graph arena in MiB. Decoder uses the maximum of prefill and decode arena settings.",
        "required": false
      },
      {
        "name": "text_decoder_weight_context_mb",
        "type": "int",
        "min": 1,
        "default": 128,
        "description": "Decoder weight context in MiB.",
        "required": false
      },
      {
        "name": "stream_decode_cache_steps",
        "type": "int",
        "default": 1024,
        "description": "Streaming decoder cache capacity in 80 ms steps. Must be positive and no greater than checkpoint sliding_window.",
        "required": false
      },
      {
        "name": "stream_batch_tokens",
        "type": "int",
        "min": 1,
        "default": 1,
        "description": "Audio tokens per streaming encoder forward.",
        "required": false
      }
    ],
    "load": []
  },
  "capabilities": {
    "asr": [
      "partial_results"
    ]
  },
  "runtime": {
    "tags": [
      "gguf",
      "stream"
    ]
  },
  "ui": {
    "recommended_package": "voxtral_realtime_q8_0",
    "tags": [
      "ASR",
      "GGUF",
      "Stream"
    ],
    "docs": [
      "docs/models/voxtral_realtime.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": "voxtral_realtime_q8_0",
      "display_name": "Voxtral Mini 4B Realtime Q8_0 GGUF",
      "default": true,
      "format": "gguf",
      "precision": "q8_0",
      "target_directory": "Voxtral-Mini-4B-Realtime-2602-GGUF",
      "files": [
        "Voxtral-Mini-4B-Realtime-2602-GGUF/voxtral-mini-4b-realtime-2602-q8_0.gguf"
      ],
      "strip_prefix": "Voxtral-Mini-4B-Realtime-2602-GGUF"
    },
    {
      "id": "voxtral_realtime_q4_k",
      "display_name": "Voxtral Mini 4B Realtime Q4_K GGUF",
      "format": "gguf",
      "precision": "q4_k",
      "target_directory": "Voxtral-Mini-4B-Realtime-2602-GGUF",
      "files": [
        "Voxtral-Mini-4B-Realtime-2602-GGUF/voxtral-mini-4b-realtime-2602-q4_k.gguf"
      ],
      "strip_prefix": "Voxtral-Mini-4B-Realtime-2602-GGUF"
    },
    {
      "id": "voxtral_realtime_bf16",
      "display_name": "Voxtral Mini 4B Realtime BF16 GGUF",
      "format": "gguf",
      "precision": "bf16",
      "target_directory": "Voxtral-Mini-4B-Realtime-2602-GGUF",
      "files": [
        "Voxtral-Mini-4B-Realtime-2602-GGUF/voxtral-mini-4b-realtime-2602-bf16.gguf"
      ],
      "strip_prefix": "Voxtral-Mini-4B-Realtime-2602-GGUF"
    }
  ],
  "sources": [
    {
      "format": "gguf",
      "roots": {
        "model": ".",
        "weights": "$gguf"
      },
      "files": {
        "config": "model:config.json",
        "generation_config": "model:generation_config.json",
        "processor_config": "model:processor_config.json",
        "tekken": "model:tekken.json"
      },
      "optional_files": {
        "params": "model:params.json",
        "readme": "model:README.md"
      },
      "tensors": {
        "weights": "weights:"
      }
    },
    {
      "format": "safetensors",
      "roots": {
        "model": "."
      },
      "files": {
        "config": "model:config.json",
        "generation_config": "model:generation_config.json",
        "processor_config": "model:processor_config.json",
        "tekken": "model:tekken.json"
      },
      "optional_files": {
        "params": "model:params.json",
        "readme": "model:README.md"
      },
      "tensors": {
        "weights": "model:model.safetensors"
      }
    }
  ]
}