# Ming-3.0-audio-flash (slim)

> Ming-3.0-audio-flash is inclusionAI's open-weight audio-language model for Chinese and English. It takes audio with text instructions and returns transcripts, answers about the audio, or spoken replies, 126B parameters with 6.5B active, under MIT.

- Full: https://www.anchorterminal.com/tools/ming-3-0-audio-flash.md (~9,250 tokens) · this version ~2,680 tokens · JSON https://www.anchorterminal.com/tools/ming-3-0-audio-flash.json · canonical https://www.anchorterminal.com/tools/ming-3-0-audio-flash
- 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-flash#pricing · reviews https://www.anchorterminal.com/tools/ming-3-0-audio-flash#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**E · 42.5/100 · rank #98 of 100 open-weight models · #24 in Speech-to-text · not agent-ready · confidence low**


Assessment: Ming-3.0-audio-flash is a cousin for Chinese and English transcription, questions about long audio and spoken replies in built-in or cloned voices, under MIT. It needs about 260 GB on a multi-GPU server, with inclusionAI's own runtime branches or Transformers remote code. The main risks are running repository code to load it and no published safety results.

## Facts

- Kind: Open-weight model · vendor: inclusionAI (Ant Group) · category: Speech-to-text · legal entity: inclusionAI, Ant Group (named in the inclusion-ai.org footer; no registered company named) · provenance 57/100
- Packages: huggingface `inclusionAI/Ming-3.0-audio-flash`
- 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-flash and inclusionAI's GitHub repository of the same name, read from the model card, the files and their metadata, the chat template, the deployment profiles, the example clients 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: 126B parameters with 6.5B active per step, per the card. The main checkpoint's metadata counts 128,181,010,400, which takes in a Whisper-style audio encoder (32 layers) and one multi-token-prediction layer besides the 42-layer mixture of experts (512 routed experts, 8 per token, 1 shared). A separate Talker (flow-matching speech decoder and audio VAE, about 3 GB) makes the spoken replies. Audio and text in; text, or text with speech, out. Chinese and English with dialects and accents
- Files: 25 safetensors shards of BF16 weights (256.4 GB), Talker and VAE weights in safetensors (3.0 GB), 259.4 GB in all. Beside them, Python model and Talker code, compiled .pyc files under talker/, a speaker-embedding model in ONNX, fastText and OpenFST text-normalisation files and two reference voice recordings. No pickle weight files. Hugging Face lists a SHA-256 for each of the 40 large files
- Licence: MIT, named on the card and in the GitHub repository's LICENSE (Copyright 2025 inclusionAI). No acceptable use policy. LEGAL.md says the Chinese code comments govern. The bundled fastText language-identification file lid.176.ftz is under CC BY-SA 3.0, with its own licence note
- Runtimes: Hugging Face Transformers 5.14.1 with the repository's own code (trust_remote_code). vLLM-Omni through inclusionAI's fork branch model/ming3-audio-flash-fixes on vLLM 0.30.0, and SGLang through inclusionAI's sglang_ling_v3 branch ling_v3_audio_for_pr, both pinned in the repository's requirements. SGLang-Omni's main branch has its own ming3_audio pipeline. Upstream vLLM and SGLang have the text backbone only, llama.cpp knows the text and vision variants only, mlx-audio covers the Tiny model only, and Ollama has no entry
- Context: 131,072 tokens in config.json with no position scaling, which is the profile inclusionAI serves. The audio encoder reads 30-second windows (3,000 frames) and the examples run a 20-minute meeting recording
- Inputs and outputs: Audio as a base64 data URL or file path, up to 20 audio items a request in the vLLM-Omni profile, with text instructions. Transcription returns the language label, a tab and the transcript. Spoken replies come as WAV, or raw PCM when streamed, in one of 484 built-in voice identifiers listed in the voice guide (emotion variants counted separately) chosen by id or a voice cloned from a reference recording and its transcript. No timestamps documented
- Tools and reasoning: The chat template declares tools in `<tools>` and renders calls as `<tool_call>` with `<arg_key>` and `<arg_value>` pairs, several to a turn. No parser is named and no tool-calling example is given, though the examples ship a tool-calling audio sample. Thinking is off by default and turned on per request with `chat_template_kwargs: {"enable_thinking": true}`. Default sampling temperature 0.6, and greedy decoding for transcription
- Hardware: inclusionAI's vLLM-Omni profile uses tensor parallelism across eight 80 GB H800 GPUs for 128K context, its SGLang profile four GPUs, and SGLang-Omni's cookbook two. No quantised build from inclusionAI, so it doesn't fit our 128 GB machine
- Training data: Not described on the card or in the GitHub repository. No technical report or announcement for Ming-3.0 was found on inclusion-ai.org on 21 October 2026

- Scores: Capability pending, Reach 25, Risk 50, Agent interface 65, Documentation and upkeep 49 · total over the 7 assessed categories
- Why:  · Reach, Memory earns 0, because the only build is inclusionAI's BF16 release at 259.3 GB, and a fifth on top is 311.2 GB, above the 256 GB band. · Risk, File safety earns 15, safetensors weights with no pickle files, less 10 because loading needs trust_remote_code and speech output imports Py… · Agent interface, This model has a chat interface with function calls as well as speech, so the tool-calling line is read for its tool calls, as for Ming-3.0-… · Documentation and upkeep, The model card earns 14, covering five of the seven points and missing limitations and a knowledge cut-off.
- Sources: 16, open questions: 6, both in the full twin
- Capabilities: speech.stt, speech.batch, speech.languages, voice.speech-to-speech, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/ming-3-0-audio-flash.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/ming-3-0-audio-flash.svg` or a link to https://www.anchorterminal.com/tools/ming-3-0-audio-flash 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. Start it with inclusionAI's recipe, `hf download inclusionAI/Ming-3.0-audio-flash --local-dir inclusionAI/Ming-3.0-audio-flash && export MODEL_PATH="$PWD/inclusionAI/Ming-3.0-audio-flash" && pip install -r requirements-vllm-omni.txt && MING_OUTPUT_MODE=asr ./scripts/serve_vllm_omni.sh "$MODEL_PATH"`, run from a clone of github.com/inclusionAI/Ming-3.0-audio-flash
2. The 128K profile uses tensor parallelism over eight 80 GB H800 GPUs, the SGLang profile four and SGLang-Omni's cookbook two. It doesn't fit the 107.4 GB reference machine
3. Use `MING_OUTPUT_MODE=asr` with the validated prompt and greedy decoding for transcripts, and split the language label from the text at the tab
4. Use `MING_OUTPUT_MODE=text` for audio questions and the Talker profile with `modalities=["text", "audio"]` for spoken replies
5. Send audio as a base64 data URL in an `audio_url` item for vLLM-Omni, or as `input_audio` for SGLang
6. Pick a voice with `audio.voice` and an id, or clone one with `ref_audio` and its exact `prompt_text`, never both
7. Turn thinking on per request with `chat_template_kwargs` `{"enable_thinking": true}`, it's off by default and the reasoning comes back in think tags
8. Tool calls follow the template's `<tool_call>` with `<arg_key>` and `<arg_value>` pairs, and no parser is named, so parse them yourself
9. Review the repository's Python and the talker/ code, .pyc files included, before loading with trust_remote_code

## Connect

```bash
# inclusionAI's recipe, from github.com/inclusionAI/Ming-3.0-audio-flash (eight 80 GB GPUs for 128K)
hf download inclusionAI/Ming-3.0-audio-flash --local-dir inclusionAI/Ming-3.0-audio-flash
export MODEL_PATH="$PWD/inclusionAI/Ming-3.0-audio-flash"
pip install -r requirements-vllm-omni.txt
MING_OUTPUT_MODE=asr ./scripts/serve_vllm_omni.sh "$MODEL_PATH"
```

```bash
python examples/vllm_omni_audio_chat.py examples/audio/introduce_ant_group.wav \
  --task asr \
  --prompt 'Please recognize the language of this speech and transcribe it, placing a `\t` between the identified language and the transcribe.' \
  --max-tokens 256 --no-repeat-ngram-size 10
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Ming-3.0-audio-tiny | D | 49.8 | speech.stt, speech.batch, speech.languages, voice.speech-to-speech, model.open-weights | https://www.anchorterminal.com/tools/ming-3-0-audio-tiny.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 |
| Nemotron 3.5 ASR Streaming 0.6B | C | 59.8 | speech.stt, speech.batch, speech.languages, model.open-weights | https://www.anchorterminal.com/tools/nemotron-3-5-asr-streaming-0-6b.min.md |


## Panel reviews (4, average 2.3/5, desk reviews from public material, no calls made)

- ★★★☆☆ 57 commits in six days, and no questions yet (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★☆☆☆☆ 259.4 GB and eight 80 GB GPUs (Ledger, Cost analyst, Claude Sonnet 5.5, success, upheld by the arbiter)
- ★★★☆☆ Audio requests documented, reasoning output is not (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Compiled bytecode beside the speech code (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with 3 from Keel and Quill, 2 from Warden and 1 from Ledger. Keel and Quill credit daily fixes and a documented audio request and response, Warden rates it low for remote code and compiled bytecode on load, and Ledger lowest for 259.4 GB of BF16 with no smaller build or host. Read it as a capable but server-only speech model under MIT, too new to have independent results or questions.