{
  "schema_version": 1,
  "family": "minimax_h3",
  "display_name": "MiniMax-H3",
  "description": "MiniMax-H3 text-to-audio/video generation through the official T2VA pipeline: Qwen3-VL text conditioning, joint audio/video DiT denoising, and MiniMax-H3 audio/video VAE decode.",
  "category": "audio_generation",
  "status": "experimental",
  "tasks": [
    "tts",
    "music",
    "sfx"
  ],
  "modes": [
    "offline"
  ],
  "languages": [
    "auto"
  ],
  "capabilities": {
    "tts": [
      "style_control",
      "multi_speaker"
    ],
    "music": [
      "instrumental",
      "style_control"
    ],
    "sfx": [
      "prompt_generation"
    ]
  },
  "runtime": {
    "tags": [
      "cuda",
      "gguf"
    ]
  },
  "ui": {
    "recommended_package": "minimax_h3_q4_k",
    "tags": [
      "TTS",
      "Music",
      "SFX",
      "GGUF"
    ],
    "docs": [
      "docs/models/minimax_h3.md",
      "docs/reports/minimax_h3_performance.md"
    ]
  },
  "options": {
    "request": [
      {
        "name": "num_inference_steps",
        "type": "int",
        "description": "Positive joint DiT denoising step count; default 50.",
        "required": false,
        "min": 1,
        "default": 50
      },
      {
        "name": "seed",
        "type": "int",
        "description": "Unsigned 32-bit generation seed in [0, 4294967295] for deterministic initial latents; default 42.",
        "required": false,
        "min": 0,
        "max": 4294967295,
        "default": 42
      },
      {
        "name": "height",
        "type": "int",
        "description": "Target video canvas height used by the joint T2VA latent layout; omitted or nonpositive values use 768. Smaller positive sizes are rounded up to the supported geometry.",
        "required": false,
        "default": 768
      },
      {
        "name": "width",
        "type": "int",
        "description": "Target video canvas width used by the joint T2VA latent layout; omitted or nonpositive values use 1344. Smaller positive sizes are rounded up to the supported geometry.",
        "required": false,
        "default": 1344
      },
      {
        "name": "num_frames",
        "type": "int",
        "description": "Target video frame count before H3 alignment; omitted or nonpositive values use 124. MiniMax-H3 video is 24 fps; positive counts are rounded to the supported temporal geometry.",
        "required": false,
        "default": 124
      },
      {
        "name": "guidance_scale",
        "type": "float",
        "description": "Classifier-free guidance scale; 1.0 matches the baseline T2VA audio run.",
        "required": false,
        "default": 1
      },
      {
        "name": "sampler_mode",
        "type": "enum",
        "description": "Joint DiT sampler: euler, res_multistep, dpmpp_2m, or unipc.",
        "values": [
          "euler",
          "res_multistep",
          "dpmpp_2m",
          "unipc"
        ],
        "required": false,
        "default": "euler"
      },
      {
        "name": "negative_prompt",
        "type": "string",
        "description": "Negative prompt used by classifier-free guidance when guidance_scale is not 1.0.",
        "required": false,
        "default": " "
      },
      {
        "name": "flow_shift",
        "type": "float",
        "description": "Video FlowMatch scheduler shift.",
        "required": false,
        "default": 12.0
      },
      {
        "name": "audio_flow_shift",
        "type": "float",
        "description": "Audio FlowMatch scheduler shift.",
        "required": false,
        "default": 3.0
      },
      {
        "name": "return_video",
        "type": "bool",
        "description": "Decode and return RGB24 video frames as an output artifact.",
        "required": false,
        "default": false
      },
      {
        "name": "text_layerwise",
        "type": "bool",
        "description": "Run the prompt encoder with scoped layer-group weights instead of a full prompt encoder weight store.",
        "required": false,
        "default": false
      },
      {
        "name": "text_layerwise_batch",
        "type": "int",
        "description": "Prompt encoder layer-group size when text_layerwise is enabled.",
        "required": false,
        "min": 1,
        "default": 1
      },
      {
        "name": "dit_layerwise",
        "type": "bool",
        "description": "Run the DiT denoiser with scoped prelude, block-group, and final weights instead of a full DiT weight store.",
        "required": false,
        "default": false
      },
      {
        "name": "dit_layerwise_batch",
        "type": "int",
        "description": "DiT block-group size when dit_layerwise is enabled.",
        "required": false,
        "min": 1,
        "default": 1
      },
      {
        "name": "dit_mlp_chunk_tokens",
        "type": "int",
        "description": "Token chunk size for the per-token DiT MLP in layerwise high-resolution mode. Zero disables chunking.",
        "required": false,
        "min": 0,
        "default": 0
      },
      {
        "name": "dit_acceleration",
        "type": "string",
        "description": "Optional DiT acceleration mode: none, first_block_cache, or spectrum. Requires dit_layerwise=false and sampler_mode=euler; Spectrum additionally requires guidance_scale=1.",
        "required": false,
        "default": "none"
      },
      {
        "name": "first_block_cache_threshold",
        "type": "float",
        "description": "First-block cache residual-difference threshold.",
        "required": false,
        "min": 0.0,
        "default": 0.1
      },
      {
        "name": "first_block_cache_start_percent",
        "type": "float",
        "description": "Denoising progress fraction where first-block caching begins; must be below end_percent. Used to derive the sigma window when explicit sigma endpoints are omitted.",
        "required": false,
        "min": 0,
        "max": 1,
        "default": 0.1
      },
      {
        "name": "first_block_cache_end_percent",
        "type": "float",
        "description": "Denoising progress fraction where first-block caching ends; must exceed start_percent. Used to derive the sigma window when explicit sigma endpoints are omitted.",
        "required": false,
        "min": 0,
        "max": 1,
        "default": 0.95
      },
      {
        "name": "first_block_cache_start_sigma",
        "type": "float",
        "description": "Largest video sigma where first-block cache may be used. Supply together with first_block_cache_end_sigma; otherwise both endpoints are derived from the denoising progress fractions and schedule.",
        "required": false
      },
      {
        "name": "first_block_cache_end_sigma",
        "type": "float",
        "description": "Smallest video sigma where first-block cache may be used. Supply together with first_block_cache_start_sigma; otherwise both endpoints are derived from the denoising progress fractions and schedule.",
        "required": false
      },
      {
        "name": "first_block_cache_max_consecutive",
        "type": "int",
        "description": "Maximum consecutive first-block cache hits before forcing a full DiT step.",
        "required": false,
        "min": 1,
        "default": 2
      },
      {
        "name": "spectrum_warmup_steps",
        "type": "int",
        "description": "Full DiT warmup steps before spectrum forecasting may skip steps.",
        "required": false,
        "min": 1,
        "default": 1
      },
      {
        "name": "spectrum_initial_window",
        "type": "float",
        "description": "Initial spectrum forecast scheduling window.",
        "required": false,
        "min": 1.0,
        "default": 2.0
      },
      {
        "name": "spectrum_flex_window",
        "type": "float",
        "description": "Spectrum forecast window growth after actual DiT steps.",
        "required": false,
        "min": 0.0,
        "default": 0.75
      },
      {
        "name": "spectrum_degree",
        "type": "int",
        "description": "Chebyshev polynomial degree for spectrum forecasting.",
        "required": false,
        "min": 0,
        "default": 1
      },
      {
        "name": "spectrum_history_size",
        "type": "int",
        "description": "Maximum spectrum forecast history rows.",
        "required": false,
        "min": 1,
        "default": 8
      },
      {
        "name": "spectrum_ridge_lambda",
        "type": "float",
        "description": "Ridge regularization for spectrum forecasting.",
        "required": false,
        "min": 0.0,
        "default": 0.1
      }
    ],
    "session": [
      {
        "name": "weight_context_mb",
        "type": "int",
        "description": "MiniMax-H3 staged weight context size in MiB.",
        "required": false,
        "min": 1,
        "default": 512
      },
      {
        "name": "mem_saver",
        "type": "bool",
        "description": "Release staged DiT, audio VAE, and video VAE weights after their request phase instead of keeping them resident for the session. Prompt encoder weights are always request-scoped.",
        "required": false,
        "default": true
      }
    ],
    "load": []
  },
  "dependencies": [],
  "package_defaults": {
    "download": {
      "kind": "huggingface_snapshot",
      "repo": "audio-cpp/audio.cpp-gguf",
      "revision": "main",
      "gated": false
    }
  },
  "packages": [
    {
      "id": "minimax_h3_q4_k",
      "display_name": "MiniMax-H3 Q4_K GGUF",
      "default": true,
      "format": "gguf",
      "precision": "q4_k",
      "target_directory": "MiniMax-H3-Q4-GGUF",
      "files": [
        "MiniMax-H3-Q4-GGUF/configuration.json",
        "MiniMax-H3-Q4-GGUF/FL2VA/processor/chat_template.json",
        "MiniMax-H3-Q4-GGUF/FL2VA/processor/merges.txt",
        "MiniMax-H3-Q4-GGUF/FL2VA/processor/preprocessor_config.json",
        "MiniMax-H3-Q4-GGUF/FL2VA/processor/tokenizer.json",
        "MiniMax-H3-Q4-GGUF/FL2VA/processor/tokenizer_config.json",
        "MiniMax-H3-Q4-GGUF/FL2VA/processor/video_preprocessor_config.json",
        "MiniMax-H3-Q4-GGUF/FL2VA/processor/vocab.json",
        "MiniMax-H3-Q4-GGUF/text_encoder_q4_k.gguf",
        "MiniMax-H3-Q4-GGUF/dit.gguf",
        "MiniMax-H3-Q4-GGUF/audio_vae_folded_f16.gguf",
        "MiniMax-H3-Q4-GGUF/video_vae.gguf"
      ],
      "strip_prefix": "MiniMax-H3-Q4-GGUF"
    },
    {
      "id": "minimax_h3_q4_k_int8_dit",
      "display_name": "MiniMax-H3 Q4_K GGUF + INT8 ConvRot DiT",
      "default": false,
      "format": "gguf",
      "precision": "q4_k",
      "target_directory": "MiniMax-H3-Q4-GGUF",
      "files": [
        "MiniMax-H3-Q4-GGUF/configuration.json",
        "MiniMax-H3-Q4-GGUF/FL2VA/processor/chat_template.json",
        "MiniMax-H3-Q4-GGUF/FL2VA/processor/merges.txt",
        "MiniMax-H3-Q4-GGUF/FL2VA/processor/preprocessor_config.json",
        "MiniMax-H3-Q4-GGUF/FL2VA/processor/tokenizer.json",
        "MiniMax-H3-Q4-GGUF/FL2VA/processor/tokenizer_config.json",
        "MiniMax-H3-Q4-GGUF/FL2VA/processor/video_preprocessor_config.json",
        "MiniMax-H3-Q4-GGUF/FL2VA/processor/vocab.json",
        "MiniMax-H3-Q4-GGUF/text_encoder_q4_k.gguf",
        "MiniMax-H3-Q4-GGUF/dit_int8.gguf",
        "MiniMax-H3-Q4-GGUF/audio_vae_folded_f16.gguf",
        "MiniMax-H3-Q4-GGUF/video_vae.gguf"
      ],
      "strip_prefix": "MiniMax-H3-Q4-GGUF"
    }
  ],
  "sources": [
    {
      "format": "gguf",
      "roots": {
        "model": "."
      },
      "files": {
        "configuration": "model:configuration.json",
        "tokenizer_config": "model:FL2VA/processor/tokenizer_config.json",
        "tokenizer_json": "model:FL2VA/processor/tokenizer.json",
        "vocab": "model:FL2VA/processor/vocab.json",
        "merges": "model:FL2VA/processor/merges.txt",
        "preprocessor_config": "model:FL2VA/processor/preprocessor_config.json"
      },
      "tensors": {
        "text_encoder_weights": "model:text_encoder_q4_k.gguf",
        "dit_weights": "model:dit.gguf",
        "audio_vae_weights": "model:audio_vae_folded_f16.gguf",
        "video_vae_weights": "model:video_vae.gguf"
      }
    }
  ]
}