{
  "hf_id": "inclusionAI/Ming-omni-tts-0.5B",
  "meta": {
    "title": "Ming-omni-tts-0.5B",
    "slug": "ming-omni-tts-0.5b",
    "provider": "inclusionAI",
    "description": "inclusionAI's dense 0.5B-LM text-to-speech model served via vLLM-Omni with style, dialect, voice-cloning and multi-speaker controls, through the OpenAI /v1/audio/speech API (44.1 kHz mono).",
    "date_added": "2026-06-17",
    "date_updated": "2026-06-15",
    "tasks": [
      "omni"
    ],
    "related_recipes": [
      "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice",
      "openbmb/VoxCPM2",
      "OpenMOSS-Team/MOSS-TTS"
    ]
  },
  "model": {
    "model_id": "inclusionAI/Ming-omni-tts-0.5B",
    "min_vllm_version": "0.22.0",
    "architecture": "dense",
    "parameter_count": "0.5B",
    "active_parameters": "0.5B",
    "context_length": 0,
    "base_args": [
      "--omni",
      "--enforce-eager"
    ],
    "base_env": {},
    "install": {
      "pip": {
        "command": "uv venv\nsource .venv/bin/activate\nuv pip install -U vllm --torch-backend auto"
      },
      "docker": {
        "command": "docker pull vllm/vllm-openai:latest"
      },
      "extras": [
        {
          "note": "vLLM-Omni provides the TTS serving backend and the `--omni` serve path. See the installation guide for pre-built wheels or building from source.",
          "command": "uv pip install git+https://github.com/vllm-project/vllm-omni.git"
        }
      ]
    }
  },
  "omni": {
    "tasks": [
      {
        "id": "t2a",
        "label": "Text → Speech",
        "endpoint": "/v1/audio/speech",
        "description": "Ming-omni-tts via /v1/audio/speech with style/dialect/cloning controls (44.1 kHz mono)",
        "curl": "curl -X POST http://localhost:8000/v1/audio/speech \\\n  -H \"Content-Type: application/json\" \\\n  -d '{\n    \"model\": \"inclusionAI/Ming-omni-tts-0.5B\",\n    \"input\": \"我觉得社会企业同个人都有责任\",\n    \"voice\": \"default\",\n    \"ref_audio\": \"file:///path/to/prompt.wav\",\n    \"response_format\": \"wav\"\n  }' --output output.wav\n"
      }
    ]
  },
  "features": {},
  "opt_in_features": [],
  "variants": {
    "default": {
      "precision": "bf16",
      "vram_minimum_gb": 16,
      "description": "BF16 weights — 0.5B LM backbone (~1.6B total incl. audio encoder + vocoder); fits a small single GPU"
    }
  },
  "hardware_overrides": {
    "amd": {
      "extra_args": [],
      "extra_env": {
        "VLLM_ROCM_USE_AITER": "0"
      }
    }
  },
  "guide": "## Overview\n\n[Ming-omni-tts-0.5B](https://huggingface.co/inclusionAI/Ming-omni-tts-0.5B) is\ninclusionAI's dense text-to-speech model served through **vLLM-Omni** with the\nOpenAI-compatible `/v1/audio/speech` API. It exposes style, dialect, voice\ncloning, and multi-speaker controls and outputs **44.1 kHz mono** audio. The\n`0.5B` refers to the LM backbone; the full checkpoint is ~1.6B parameters\nincluding the audio encoder and vocoder.\n\n## Prerequisites\n\n- **Hardware**: a single GPU. The reference end-to-end run was on **AMD\n  `gfx942` (ROCm 7.x)** using the `vllm/vllm-omni-rocm:v0.22.0` image; the\n  NVIDIA CUDA path uses the same `vllm serve` command shown below.\n- **vLLM-Omni** targeting vLLM >= 0.22.\n- Ming's `model_type` falls through deploy-config auto-detection, so the\n  launch command passes `--deploy-config vllm_omni/deploy/ming_tts.yaml`\n  explicitly (the file ships in the vLLM-Omni repo/package).\n- The tested environment uses `--enforce-eager`.\n\n## Installation\n\n```bash\nuv venv && source .venv/bin/activate\nuv pip install git+https://github.com/vllm-project/vllm-omni.git\n```\n\n## Launch the server\n\n### NVIDIA (CUDA)\n\n```bash\nvllm serve inclusionAI/Ming-omni-tts-0.5B \\\n  --deploy-config vllm_omni/deploy/ming_tts.yaml \\\n  --omni --enforce-eager --port 8000\n```\n\n### AMD (ROCm) — reference-tested via the official image\n\n```bash\ndocker run --rm \\\n  --group-add=video --ipc=host --cap-add=SYS_PTRACE \\\n  --security-opt seccomp=unconfined \\\n  --device /dev/kfd --device /dev/dri \\\n  -v ~/.cache/huggingface:/root/.cache/huggingface \\\n  -v \"$PWD\":/app/vllm-omni -w /app/vllm-omni \\\n  -e VLLM_ROCM_USE_AITER=0 \\\n  -p 8000:8000 \\\n  vllm/vllm-omni-rocm:v0.22.0 \\\n  --model inclusionAI/Ming-omni-tts-0.5B \\\n  --deploy-config vllm_omni/deploy/ming_tts.yaml \\\n  --omni --port 8000 --enforce-eager\n```\n\n## Client usage\n\nDialect control with a reference clip (e.g. Cantonese / 广粤话). `--ref-audio`\nmatches upstream `use_spk_emb=True`; do **not** add `--ref-text` for the\ndialect case:\n\n```bash\npython examples/online_serving/text_to_speech/ming_tts/openai_speech_client.py \\\n  --text \"我觉得社会企业同个人都有责任\" \\\n  --instruction-json '{\"方言\":\"广粤话\"}' \\\n  --ref-audio /path/to/yue_prompt.wav \\\n  --max-new-tokens 200 \\\n  --output dialect.wav\n```\n\n## Known limitations\n\n- Output is **44.1 kHz mono**.\n- The reference validation in the source recipe was on AMD `gfx942` (ROCm);\n  the CUDA command above is the equivalent path.\n\n## References\n\n- [Model card](https://huggingface.co/inclusionAI/Ming-omni-tts-0.5B)\n- [Ming-omni-tts upstream repo](https://github.com/inclusionAI/Ming-omni-tts)\n- [vLLM-Omni Ming-TTS online example](https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/text_to_speech/ming_tts)\n- [vLLM-Omni Ming-TTS offline example](https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/text_to_speech/ming_tts)\n",
  "source": {
    "source": "vllm-omni",
    "repo": "vllm-project/vllm-omni",
    "ref": "main",
    "sha": "cda4a1ef83b295ba332146a7709e6277725edc9f",
    "path": "recipes/models/inclusionAI/Ming-omni-tts-0.5B.yaml",
    "edit_url": "https://github.com/vllm-project/vllm-omni/edit/main/recipes/models/inclusionAI/Ming-omni-tts-0.5B.yaml",
    "overlays": []
  }
}