# Fix list: Ming-3.0-audio-flash From Anchor Terminal's listing at https://www.anchorterminal.com/tools/ming-3-0-audio-flash, the October 2026 research run, assessed 21 October 2026. Grade E, 42.5 out of 100. This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public. For a coding agent working on Ming-3.0-audio-flash: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published. ## 1. Reach, 25 out of 100, up to 28.8 more on the total Why it scored 25: 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. Runtimes earn 0, since no named runtime runs the audio model from its own released code. Transformers needs the repository's remote code, vLLM and SGLang need inclusionAI's fork branches, vLLM-Omni and SGLang-Omni aren't on the list, and llama.cpp, Ollama and LM Studio don't have it. Builds earn 0 with BF16 only. Terms earn 15 for MIT with nothing added. Download earns 10, since the repository isn't gated. Hosted earns 0, with no inference provider on its Hugging Face page, none on OpenRouter and no inclusionAI API for it. The checklist (https://www.anchorterminal.com/benchmark/#checklist-reach): Whether an agent can run it, and where. - 0 to 30, memory. The smallest build from the publisher or a runtime's own library that keeps 4-bit precision or better, at its file size plus a fifth for the runtime and a working context. 30 for 16 GB or less, 25 for 24 GB, 20 for 32 GB, 15 for 64 GB, 10 for what fits our reference machine (the GPU share of 128 GB of unified memory), 5 for 256 GB, 0 above. - 0 to 25, runtimes. 5 for each of llama.cpp, Ollama, vLLM, SGLang, LM Studio and Hugging Face Transformers that runs the architecture in a released version, from the runtime's own code, library or documentation. A publisher's fork or an unreleased branch doesn't count. 25 at most. - 0 to 15, builds from the publisher. 15 for quantised builds at two precisions or more (FP8, 4-bit, GGUF), 10 for one, 0 when only other people's quantisations exist. - 0 to 15, terms of use. 15 for an OSI-approved licence with nothing added, 10 for a custom licence that lets anyone use it commercially for free below a threshold or under a use policy, 5 for research or non-commercial use only, 0 for no licence or unclear terms. - 0 to 10, download. 10 when the files download with no account and no terms to accept, 5 for a click-through that needs an account, 0 when a person approves each request. - 5, hosted too. The publisher's own API serves it, or at least two inference providers do, so an agent without the hardware can still call it. When our reference-machine runs start, speed on that machine (tokens per second and time to first token at the build that fits) joins Reach. ## 2. Risk, 50 out of 100, up to 15.4 more on the total Why it scored 50: File safety earns 15, safetensors weights with no pickle files, less 10 because loading needs trust_remote_code and speech output imports Python from the model directory, which also ships compiled .pyc files. Provenance earns 15, files in inclusionAI's own organisation, linked from its site, and a checksum on every file. Safety evaluation earns 0, with none published. Agent guidance earns 0, since the only caution is about reviewing code before loading it. Licence risk keeps 15 under plain MIT. The security programme earns 5 for Ant Group's AntSRC reporting centre, read as a company-wide route the way Google's was for Gemma, while advisories handled in public weren't found or couldn't be checked on GitHub. The checklist (https://www.anchorterminal.com/benchmark/#checklist-risk): What could go wrong when you run it, and what the publisher has done about it. A higher score means less risk. - 0 to 25, file safety. 25 for weights in safetensors or GGUF only, 15 when pickle files (.bin, .pt, .pth) sit beside them, 0 for pickle only. 10 off when the documented way to load it or to build its prompts runs code from the repository (`trust_remote_code`, or the publisher's scripts in place of a template). - 0 to 15, provenance. 10 when the files come from the publisher's own organisation, the one its site or docs link to, and 5 for a checksum on every file. - 0 to 20, safety evaluation. 20 when results are published and someone other than the publisher ran them or can rerun them, 10 for the publisher's own results, 0 for none. - 0 to 15, guidance for agent use. 15 for documented prompt-injection guidance or an instruction hierarchy the model was trained to follow, 10 for a statement of what its trained refusals cover and the safeguards to add around it, 5 for general cautions, 0 for none. - 0 to 15, licence risk. 15 to start, 5 off for each of these, a use policy or field-of-use limits beyond the law, terms the licensor can change or revoke, a revenue or user threshold, and a licence that changed without a dated note or whose link is broken. Not below 0. - 0 to 10, a security programme. 5 for a way to report a flaw (security.txt, a disclosure policy or a security contact), 5 for advisories handled in public. When our reference-machine runs start, measured robustness (prompt injection with AgentDojo, over-refusal with XSTest) joins Risk. ## 3. Documentation and upkeep, 49 out of 100, up to 7.8 more on the total Why it scored 49: The model card earns 14, covering five of the seven points and missing limitations and a knowledge cut-off. Training data earns 0, with nothing described. Last change earns 20, with the voice interface, examples and inclusionAI's runtime branch all changed between 18 and 20 October 2026. Responsiveness earns 0, because nobody has asked a question yet in the Hugging Face discussions and GitHub issues couldn't be counted. Version policy earns 5 for a first release in its own repository. File health earns 10, with config, tokenizer, template and card in agreement and no broken-file report waiting. The checklist (https://www.anchorterminal.com/benchmark/#checklist-upkeep): - 0 to 20, the model card. Intended use, limitations, architecture, parameters, context length, languages and knowledge cut-off. 20 for all seven, 3 off for each that's missing. - 0 to 20, training data. 20 for the data published or listed with sources and sizes, 15 for a data card that names the sources or the mix, the cut-off and the filtering, 10 for a general description with a cut-off, 5 without one, 0 for none. - 0 to 20, the last change to the files, the template or the reference code. 20 within 30 days, 15 within 90, 10 within 180, 5 within a year, 0 older. - 0 to 20, responsiveness. The publisher's replies to the 40 most recent questions and reports in the repository's discussions or issues. 20 for most, 10 for some, 5 for a few, 0 for none. - 0 to 10, version policy. 10 for earlier versions kept in their own repositories with a dated note when one is replaced, 5 for earlier versions kept without notes or a first release, 0 when files were replaced in place without a note. - 0 to 10, file health. 10 when config, tokenizer, template and card agree and no report of a broken file waits without a reply, 5 when one does, 0 when the files contradict the card. ## 4. Agent interface, 65 out of 100, up to 5.4 more on the total Why it scored 65: 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-audio-tiny, and structured output for what it returns. The chat template earns 20, a Jinja file in the repository. The tool-calling format earns 10, because function tools are declared and returned only as the template says, with no parser named, though inclusionAI's README documents the audio requests and replies for its vLLM-Omni and SGLang servers. Trained tool calling earns 20 from inclusionAI's own tool-calling sample and a template that renders several calls in one turn. Structured output earns 5, since the transcript form and audio format are documented but no timestamps or JSON output are. Reasoning control earns 0, because thinking switches per request but where its text comes back isn't documented. Context earns 10 for 131,072 tokens with no scaling. The checklist (https://www.anchorterminal.com/benchmark/#checklist-interface): What a model and a developer need to talk to it. Speech, image and other models with no chat interface read the tool-calling and structured-output lines for what the model takes and returns (voices, timestamps, streaming, typed outputs), and the note says how. - 0 to 20, the chat template. 20 for a template in the repository (Jinja, in the tokenizer config or its own file), 10 when the prompt format is only given as code you run. - 0 to 25, the tool-calling format. 25 when the publisher documents how tools are declared and how calls come back, with the parser to use in a named runtime, 10 when only the template or a runtime's own recipe says it. - 0 to 20, tool calling the model was trained for. 20 with several calls in one turn, 10 for one call at a time. - 0 to 15, structured output. 15 when JSON or JSON Schema output is documented for the model with how to ask for it, 5 when the card claims it without saying how. - 10, reasoning control. Thinking that can be turned off or budgeted per request, with where the reasoning text comes back. - 0 to 10, context. 10 for 128K tokens or more trained, 5 for 32K to 127K. Context reached only by scaling the position encoding counts one band down. ## What we couldn't check What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it. - unchecked, GitHub issues and security advisories on inclusionAI/Ming-3.0-audio-flash, since GitHub's API and site weren't open to this session, so responsiveness and public advisories score as absent - Whether AntSRC's scope covers inclusionAI's open models wasn't checked, since the site is a script-rendered app. The 5 rests on it being Ant Group's company-wide reporting route, read the same way for all five Ming-3.0 listings - The tool-calling format is read for the model's text tool calls (10), as for Ming-3.0-audio-tiny, since the checklist reads a speech model for what it takes and returns only when it has no chat interface. Read purely as a speech model, both audio listings would earn 25 - Trained tool calling rests on inclusionAI's tool_calling.wav sample and the template. The card doesn't say the model was trained for tool calls - Capability is pending unless independent results (the Open ASR Leaderboard, Seed-TTS-eval runs) appear; inclusionAI's own WER and CER table doesn't count - The card's 126B total doesn't say what it counts; the checkpoint metadata gives 128,181,010,400 with the audio encoder and one prediction layer, plus about 3 GB of Talker and VAE ## Weaknesses - 259.3 GB of BF16 files and no quantised build, so it needs a multi-GPU server and doesn't fit 128 GB - No named runtime runs it from released code, it needs inclusionAI's fork branches or trust_remote_code - No hosted provider, so an agent without the hardware can't call it - No safety results, training data, limitations or knowledge cut-off published - No timestamps or JSON transcript output, and where reasoning text returns isn't documented ## What costs an agent a turn today The notes we give agents before they call it. Each one is a workaround an agent shouldn't need. - 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 - 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 - 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 - Use `MING_OUTPUT_MODE=text` for audio questions and the Talker profile with `modalities=["text", "audio"]` for spoken replies - Send audio as a base64 data URL in an `audio_url` item for vLLM-Omni, or as `input_audio` for SGLang - Pick a voice with `audio.voice` and an id, or clone one with `ref_audio` and its exact `prompt_text`, never both - 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 - Tool calls follow the template's `` with `` and `` pairs, and no parser is named, so parse them yourself - Review the repository's Python and the talker/ code, .pyc files included, before loading with trust_remote_code ## What the review panel asked for - tagged runtime release - named file versions - an FP8 or 4-bit build - a hosted endpoint - word timestamps - remove the compiled .pyc files - native runtime support ## When it's done Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.