# Ming-3.0-audio-tiny (slim)

> Ming-3.0-audio-tiny is inclusionAI's open-weight audio-language model for Chinese and English speech recognition, audio understanding and spoken replies, 9.3B parameters with 2.7B active, published on Hugging Face under MIT.

- Full: https://www.anchorterminal.com/tools/ming-3-0-audio-tiny.md (~9,050 tokens) · this version ~2,330 tokens · JSON https://www.anchorterminal.com/tools/ming-3-0-audio-tiny.json · canonical https://www.anchorterminal.com/tools/ming-3-0-audio-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-audio-tiny#pricing · reviews https://www.anchorterminal.com/tools/ming-3-0-audio-tiny#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**D · 49.8/100 · rank #95 of 100 open-weight models · #22 in Speech-to-text · not agent-ready · confidence low**


Assessment: A small MIT-licensed audio-language model that transcribes Chinese and English, answers questions about audio and replies in a built-in or cloned voice. It needs about 25 GB and SGLang-Omni from main or inclusionAI's own vLLM-Omni and SGLang branches. The main risk is that every loading path runs repository code, with nothing published on safety.

## Facts

- Kind: Open-weight model · vendor: inclusionAI (Ant Group) · category: Speech-to-text · legal entity: not named · provenance 31/100
- Packages: huggingface `inclusionAI/Ming-3.0-audio-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-audio-tiny, read from the model card, the files and their metadata, the chat template, inclusionAI's GitHub repository Ming-3.0-audio-flash (which covers both sizes) 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: 9.3B parameters with 2.7B active per step, per the card. The Hub's metadata counts 8,871,306,016 in the main shards, which leaves out the Talker and its VAE (1.40 GB and 1.56 GB files). A Whisper-style audio encoder (32 layers) and projector feed a 24-layer mixture-of-experts backbone with 128 routed experts and 1 shared, 8 routed per token. An optional Talker turns the answer into speech
- Inputs and outputs: Audio plus text instructions in, up to 20 audio clips per prompt in the vLLM-Omni profile. Text out, or text and speech. Transcripts come back as a language label, a tab and the text. Speech in about 480 built-in voice identifiers (including emotion variants) or a cloned voice from a reference WAV and its transcript, streamed through vLLM-Omni. No timestamps documented
- Languages: Chinese and English, including accents and Chinese dialects, per the card
- Files: 130 files, 20.78 GB. Weights in safetensors only (two thinker shards, the Talker and its VAE), plus an ONNX speaker model, a fastText language-ID model, text-normalisation FSTs, Python model code and 40 compiled .pyc files in __pycache__ folders under talker/. No pickle weights. Hugging Face lists a SHA-256 for every large file. No LICENSE file in the repository; the card links MIT on GitHub
- Licence: MIT (Copyright 2025 inclusionAI) per the card and the GitHub LICENSE. The bundled lid.176.ftz is fastText's model under CC BY-SA 3.0, with its notice in talker/data. LEGAL.md says the Chinese code comments govern
- Runtimes: vLLM-Omni through inclusionAI's model/ming3-audio-flash-fixes branch on vLLM 0.30.0, SGLang through inclusionAI's sglang_ling_v3 audio branch (text output), and SGLang-Omni's main branch, which serves transcription, chat and speech. Transformers 5.14.1 with trust_remote_code. mlx-audio's main branch runs the community MLX builds. vLLM's main registry, llama.cpp (text backbone only) and Ollama don't run the audio model
- Context: 131,072 positions in config.json and the vLLM-Omni profiles. The card doesn't state a context length. inclusionAI's SGLang script defaults Tiny to 8,192
- Tools and reasoning: The chat template declares tools and renders one or more calls per turn as `<tool_call>name` with `<arg_key>` and `<arg_value>` pairs. No parser is named. Thinking is off by default and turned on per request with `chat_template_kwargs: {"enable_thinking": true}`. Recommended sampling `temperature=0.6`, `top_p=1.0`; ASR uses greedy decoding with 10-gram blocking
- Hardware: One GPU for text or ASR, and one GPU per stage (thinker and Talker) in the reference speech profile, or one GPU for both in SGLang-Omni with `--talker.gpu 0`. No memory figure is stated. The BF16 files total 20.78 GB

- Scores: Capability pending, Reach 45, Risk 50, Agent interface 65, Documentation and upkeep 46 · total over the 7 assessed categories
- Why:  · Reach, Memory 20, the only build inclusionAI publishes is BF16 at 20.78 GB, which with a fifth on top is 24.9 GB, just past the 24 GB band, so it l… · Risk, File safety 15, safetensors only with no pickle weights, less 10 because every documented way to load it runs the repository's code, and 40… · Agent interface, Read for open weights and, as a speech model, for what it takes and returns, while this model also has a chat interface with function calls. · Documentation and upkeep, Model card 11, it covers intended use, architecture, parameters and languages and leaves out limitations, context length and a knowledge cut…
- Sources: 18, open questions: 6, both in the full twin
- Capabilities: speech.stt, speech.batch, speech.languages, speech.voices, voice.speech-to-speech, voice.tools, voice.clone, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/ming-3-0-audio-tiny.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/ming-3-0-audio-tiny.svg` or a link to https://www.anchorterminal.com/tools/ming-3-0-audio-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. Fastest start is SGLang-Omni from main, `sgl-omni serve --model-path inclusionAI/Ming-3.0-audio-tiny --talker.gpu 0`, which exposes OpenAI-style /v1/audio/transcriptions, /v1/audio/speech and /v1/chat/completions
2. For transcripts use greedy decoding with 10-gram blocking and the validated ASR prompt, and expect a language label, a tab, then the text
3. Parse tool calls yourself from `<tool_call>name` with `<arg_key>` and `<arg_value>` pairs, since no runtime parser is named
4. Thinking is off by default, turn it on with chat_template_kwargs enable_thinking true and strip the `<think>` block from the reply yourself
5. In vLLM-Omni pick a voice with audio.voice id from docs/voices.md, and don't combine it with a reference recording
6. Review the repository's Python before passing trust_remote_code, and consider deleting the __pycache__ folders under talker/ so only readable source runs

## Connect

```bash
# SGLang-Omni (main branch has the Ming-3.0 audio pipeline); install it per https://sgl-project.github.io/sglang-omni/
sgl-omni serve --model-path inclusionAI/Ming-3.0-audio-tiny --talker.gpu 0
```

```bash
curl http://127.0.0.1:8000/v1/audio/transcriptions \
  -F model=inclusionAI/Ming-3.0-audio-tiny \
  -F file=@input.wav \
  -F temperature=0
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Omni-30B-A3B-Instruct | E | 41.8 | speech.stt, speech.languages, speech.voices, voice.speech-to-speech, voice.tools, model.open-weights | https://www.anchorterminal.com/tools/qwen3-omni-30b-a3b-instruct.min.md |
| Ming-3.0-audio-flash | E | 42.5 | speech.stt, speech.batch, speech.languages, voice.speech-to-speech, model.open-weights | https://www.anchorterminal.com/tools/ming-3-0-audio-flash.min.md |
| Parakeet TDT 0.6B v3 | B | 66.9 | speech.stt, speech.batch, speech.languages, model.open-weights | https://www.anchorterminal.com/tools/parakeet-tdt-0-6b-v3.min.md |
| Qwen3-ASR-1.7B | B | 64.4 | speech.stt, speech.batch, speech.languages, model.open-weights | https://www.anchorterminal.com/tools/qwen3-asr-1-7b.min.md |
| Qwen3-ASR-0.6B | C | 61.2 | speech.stt, speech.batch, speech.languages, model.open-weights | https://www.anchorterminal.com/tools/qwen3-asr-0-6b.min.md |


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

- ★★★☆☆ Daily commits to the serving profiles (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★☆☆ 20.78 GB, one GPU per stage (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Function calls with no parser and no example (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Every load path runs the publisher's code (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 Keel and Ledger and 2 from Warden and Quill. They agree the files are consistent and freshly maintained but that every load path runs inclusionAI's code, and they split on whether a single-GPU route and daily fixes outweigh tool calls defined only by the template and no safety work. Read it as a small, new speech model for a 32 GB machine that needs its own code reviewed before use.