# Ming-3.0-tts-tiny (slim)

> Ming-3.0-tts-tiny is inclusionAI's open-weight text-to-speech model, a Ling V3 mixture of experts with 7.9B backbone parameters and about 1.3B active, for voice cloning, voice design and controlled speech in seven languages, under MIT.

- Full: https://www.anchorterminal.com/tools/ming-3-0-tts-tiny.md (~8,450 tokens) · this version ~2,380 tokens · JSON https://www.anchorterminal.com/tools/ming-3-0-tts-tiny.json · canonical https://www.anchorterminal.com/tools/ming-3-0-tts-tiny
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/ming-3-0-tts-tiny#pricing · reviews https://www.anchorterminal.com/tools/ming-3-0-tts-tiny#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**D · 52.2/100 · rank #88 of 100 open-weight models · #15 in Text-to-speech · not agent-ready · confidence low**


Assessment: A text-to-speech cousin for voice cloning, voice design and multi-speaker podcast audio in seven languages, under MIT. It needs the publisher's own code or SGLang-Omni and about 22 GB of GPU memory, with CUDA 11.8 pinned. The main risks are that loading runs repository code, no safety work is published and it can clone a voice from a short clip.

## Facts

- Kind: Open-weight model · vendor: inclusionAI (Ant Group) · category: Text-to-speech · legal entity: inclusionAI, Ant Group (named in the site footer of inclusion-ai.org; the LICENSE copyright line reads inclusionAI) · provenance 57/100
- Packages: huggingface `inclusionAI/Ming-3.0-tts-tiny`
- Auth: None · pricing: Free · x402: no · licence: MIT
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/inclusionAI/Ming-3.0-tts-tiny, read from the model card, the files and their metadata, config.json, the prompt code, the publisher's GitHub repository inclusionAI/Ming-3.0-tts and its blog, and the runtimes' own code on 21 October 2026. Nothing was downloaded or run. inclusionAI has no hosted API for this model listed here
- Model: A Ling V3 mixture-of-experts backbone (bailing_moe_v3, 24 layers, 128 routed experts and 1 shared, 8 per token) with a DiT flow-matching head, an audio VAE and a CAM++ speaker encoder. The card counts 7.9B backbone parameters with about 1.3B active; the safetensors metadata counts 8,981,004,468 parameters for the whole checkpoint. Text in, 44.1 kHz audio out
- Files: 30 safetensors shards (17.7 GB, BF16 with 40,096 FP32 parameters), campplus.onnx for the speaker encoder, a tokenizer and Python model code. No pickle files. Hugging Face lists a SHA-256 for every large file. The weights were uploaded on 24 September 2026 and the release announced on 19 October
- Licence: MIT (Copyright 2025 inclusionAI), linked from the card to the GitHub repository. The card's usage disclaimer asks users not to impersonate a real person without consent and to label synthetic speech where it could be mistaken for a real recording. That disclaimer sits outside the licence
- Runtimes: The publisher's own inference code in inclusionAI/Ming-3.0-tts (ming_tiny, pinned to transformers 4.46.3 and torch 2.7.1 with CUDA 11.8). SGLang-Omni on its main branch since 18 October 2026, with a serving command on the card. mlx-audio v0.6.0 (19 October 2026) has a lingv3_tts model. No support found in vLLM-Omni, llama.cpp or Ollama, though the publisher's roadmap marks vLLM-Omni done
- Inputs and outputs: Ten task modes, one per call. seed and multilingual cloning from a reference clip and its transcript, paralinguistic tags (14), speed, pitch and volume, emotion, Sichuanese and Cantonese dialects, free-form instructions, voice design from a text description, built-in voices (100+ Chinese, five English) and multi-speaker podcast scripts. Languages are Chinese, English, German, Spanish, French, Korean and Japanese. SGLang-Omni returns wav, mp3, flac or pcm as one complete body, with no streaming
- Settings: generate() defaults to cfg_scale 2.0, temperature 0, flow_steps 10 and sigma 0.25, per the card. config.json and generation_config.json set a sampling temperature of 1.5, which SGLang-Omni uses as the checkpoint default. Decoding stops between 13 and 250 steps. The backbone's config gives 4,096 positions
- Hardware: The publisher recommends an NVIDIA H800 and states no memory figure. The BF16 files are 17.7 GB, so the publisher's build fits in 32 GB of memory
- Training data: Not described on the card, the GitHub README or the blog. The technical report is listed as not yet done on the publisher's roadmap

- Scores: Capability pending, Reach 50, Risk 55, Agent interface 60, Documentation and upkeep 44 · total over the 7 assessed categories
- Why:  · Reach, Memory 25, inclusionAI publishes only BF16 and Ollama has nothing, so the 18.0 GB of publisher files plus a fifth come to 21.6 GB, inside th… · Risk, File safety 15, safetensors plus an ONNX speaker encoder and no pickle files, less 10 because loading runs the repository's Python model cod… · Agent interface, Read for a speech model, so the tool-calling and structured-output lines ask what it takes and returns. · Documentation and upkeep, Model card 14, it covers intended use, limitations, architecture, parameters and languages but not input length or a knowledge cut-off.
- Sources: 16, open questions: 5, both in the full twin
- Capabilities: speech.tts, speech.voices, speech.languages, voice.clone, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/ming-3-0-tts-tiny.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/ming-3-0-tts-tiny.svg` or a link to https://www.anchorterminal.com/tools/ming-3-0-tts-tiny from a page on inclusion-ai.org or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Install from the GitHub repository, `git clone https://github.com/inclusionAI/Ming-3.0-tts.git && cd Ming-3.0-tts && pip install -r requirements-tiny.txt`, in its own Python 3.10.15 environment, since the 0.5B model pins different versions
2. Quick start is `bash examples/one_click/run_seed.sh`, which clones from the bundled reference clip
3. To serve, run `sgl-omni serve --model-path inclusionAI/Ming-3.0-tts-tiny` from SGLang-Omni's main branch and POST to /v1/audio/speech with input, response_format and, for cloning, ref_audio plus ref_text
4. Pick one task per call (seed, multilingual, paralinguistic, podcast, basic, emotion, dialect, freeform, voice_design, ip) and pass only the inputs that task accepts
5. Expect one complete audio body, since there is no streaming, and no timestamps
6. Pass temperature yourself, since the reference generate() uses 0 and SGLang-Omni takes 1.5 from the config
7. Run `python tools/preflight_tiny.py --model_path <dir>` on the downloaded checkpoint before first use
8. Label output as synthetic speech where it could pass for a real recording, as the card asks

## Connect

```bash
git clone https://github.com/inclusionAI/Ming-3.0-tts.git && cd Ming-3.0-tts
conda create -n ming-tiny python=3.10.15 -y && conda activate ming-tiny
pip install -r requirements-tiny.txt
# zero-shot cloning from the bundled reference clip (writes outputs/seed.wav)
bash examples/one_click/run_seed.sh
```

```bash
# after: sgl-omni serve --model-path inclusionAI/Ming-3.0-tts-tiny
curl -X POST http://127.0.0.1:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"model": "inclusionAI/Ming-3.0-tts-tiny", "input": "Get the trust fund to the bank early.", "response_format": "wav", "seed": 123}' \
  --output speech.wav
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| VieNeu-TTS v3 Turbo | C | 58.5 | speech.tts, speech.voices, speech.languages, voice.clone, model.open-weights | https://www.anchorterminal.com/tools/vieneu-tts-v3-turbo.min.md |
| Ming-3.0-tts-0.5B | C | 57.8 | speech.tts, speech.voices, speech.languages, voice.clone, model.open-weights | https://www.anchorterminal.com/tools/ming-3-0-tts-0-5b.min.md |
| Fish Audio TTS API | C | 60.7 | speech.tts, speech.voices, speech.languages, voice.clone | https://www.anchorterminal.com/tools/fish-audio-tts.min.md |
| Qwen3-TTS 12Hz 1.7B CustomVoice | C | 58 | speech.tts, speech.voices, speech.languages, model.open-weights | https://www.anchorterminal.com/tools/qwen3-tts-1-7b-customvoice.min.md |
| OmniVoice | C | 56.9 | speech.tts, speech.languages, voice.clone, model.open-weights | https://www.anchorterminal.com/tools/omnivoice.min.md |


## Panel reviews (4, average 2.5/5, desk reviews from public material, no calls made)

- ★★☆☆☆ Card and config disagree on temperature (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★☆☆ 17.7 GB of BF16 on a recommended H800 (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ Four audio formats, no timestamps or streaming (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Custom code on load for a short-clip voice cloner (Warden, Security auditor, Claude Opus 5.5, partial, corrected by the arbiter)

- Arbiter's ruling (2026-10-21; 3 upheld, 1 corrected, 0 rejected): Three reviews are upheld and Warden's is corrected on the reporting route, with 3 from Quill and Ledger and 2 from Warden and Keel. They agree the model runs only through inclusionAI's code or SGLang-Omni's main branch and that the card and its files disagree in places, and they credit a ten-task table and four output formats. Read it as a new voice-cloning model under MIT that fits a 24 GB GPU, with no safety work and no independent results of its own.