Ming-3.0-audio-flash Shipped Six Weeks Before Its Own Model Card

A 126-billion-parameter audio-language model has been downloadable from Hugging Face since 3 September 2026, and it has 22 downloads and eight likes to show for it. Until this week the only description attached to the repository was a stale card that called itself "Ming Audio V3" and said the repo "currently documents the ASR path." Ming-3.0-audio-flash is the name it carries now. The weights are not new. What arrived across 13–19 October 2026 is everything that makes a checkpoint usable by someone who did not train it: a config that finally claims a 128K context, a re-upload that turns a video-language-audio model into an audio-only one on a Transformers 5 code path, a voice pack with two reference clips, and the first real model card the repository has ever had. inclusionAI announced none of it. No third party has published a single figure for the model, so everything below is either read directly off the commit history or quoted from the vendor's own evaluation images — which is why it is worth saying at the top that the model is from September and only the documentation is from this week. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

What changed, and when

The commit log is unusually legible, and it tells a story the model card does not. The repository was created on 3 September 2026 at 11:49 UTC with one file in it, a .gitattributes. Nine minutes later 25 safetensors shards totalling 257 GB landed under a commit titled "Add files using upload-large-folder tool" — the standard signature of a training team pushing a checkpoint, not a product team staging a launch. The accompanying README declared Apache 2.0, tagged the repo automatic-speech-recognition, and told you to instantiate the model with trust_remote_code=False.

Then the repository went quiet for six weeks. When it moved again, it moved four times in seven days.

• 13 October, 12:17 UTC — "Upload config.json." One file, and the interesting one: max_position_embeddings goes from 8,192 to 131,072. Nothing else in the repository changes, so for two days the repo advertised a 128K context with no modelling code able to use it.

• 15 October, 09:35 and 09:40 UTC — "Remove deprecated files" followed by "Upload updated model." The weight layout did not change, but 46 files were rewritten and the first shard came down from 10.64 GB to 9.77 GB. The config's audio model_type changed from bailing_moe_v3_vl_audio to bailing_moe_v3_audio, a leftover video_patch_token was deleted, the declared transformers_version jumped from 4.56.2 to 5.14.1, and auto_map gained AutoModel and AutoProcessor entries that were not there before. In the same push, the README and the examples/ folder disappeared — for the next four days the repository had no model card at all.

• 17 October, 11:44 UTC — "Add voice presets." A voice_presets/ folder appears with US_James.wav, 灵楷.wav and a voice_prompts.json holding their transcripts.

• 19 October, 11:35 UTC onwards — "Add model card," then assets, then a dozen small commits polishing navigation and link alignment, ending at 13:20 UTC with "Add Tiny model link." This is the commit that renamed the model, changed the licence metadata to MIT, and added the 8,259-byte card that describes a full audio-language model rather than an ASR path. The README it replaced had been deleted four days earlier; the version before that, from September, was titled "Ming Audio V3" and declared Apache 2.0.

There is a second repository worth knowing about. Ming-3.0-audio-tiny was created at the same moment as the flash model, on 3 September at 11:45 UTC, and got the same treatment on 19 October. It is the smaller sibling: 9.3B total parameters with 2.7B active, MIT, roughly 21 GB across four shards, and three downloads.

What you can actually download

The configuration is published, so the architecture is not a mystery. What changed this week is that the config and the code finally agree. Custom modelling files have been in the repository since the first upload — modeling_bailing_moe_v3_audio.py and processing_bailing_moe_v3_audio.py were both there on 3 September — but the config they were supposed to load described a video-language-audio model with an 8K window, and the card told you to instantiate it with trust_remote_code=False, which is the opposite of what custom code needs. The 15 October re-upload is the version where audio_config.model_type reads bailing_moe_v3_audio, the video patch token is gone, the declared runtime is transformers_version 5.14.1, and auto_map registers AutoModel and AutoProcessor alongside the two entries that were always there.

• Size — 25 bfloat16 safetensors shards totalling 256 GB for the language model, plus a 3 GB Talker and VAE pair under talker/. The tiny model is about 21 GB in four shards.

• Backbone — 42 layers, hidden size 2,560, 512 experts with 8 active per token, vocabulary 157,184. That is 126B total and 6.5B active per inference step, per the card.

• Audio tower — a Whisper-style encoder at 32 layers and width 1,280 with 20 heads, 128 mel bins, and a 3,000-frame audio context. The audio token id is 157,168.

• Context — 131,072 positions. It was 8,192 in the 3 September config, and the 13 October config that first raised it still called itself bailing_moe_v3_vl_audio and still carried the video patch token; the clean audio-only version is the 15 October one.

• Output — text, or text plus synthesised speech through an optional Talker. The card names a default IP voice, 灵小柔, and two preset reference voices, 灵楷 and US_James, whose audio and transcripts are both in the repo.

• Licence — MIT. It was Apache 2.0 until 19 October. Note that the repository itself contains no LICENSE file: the MIT claim lives in the card's metadata and in a link to a GitHub path, and the only licence text actually present in the repo belongs to a bundled language-identification model under talker/data/.

The quick start does not resolve yet

This is the part a reader planning to evaluate the model needs before anything else, and it is not in the card. The card's install block clones github.com/inclusionAI/Ming-3.0-audio-flash. That URL returns HTTP 404 today, from both the web and the API. So does every path the card builds on top of it: examples/transformers_audio.py, examples/transformers_thinker_talker.py, docs/voices.md, scripts/serve_vllm_omni.sh, scripts/serve_sglang.sh, and requirements-transformers.txt.

None of those exist in the Hugging Face repository either. It carries no examples/ directory — an earlier one was deleted on 17 September — no scripts/, no docs/, and no requirements file outside talker/. What it does carry is everything needed to load the model with Transformers directly, which the card also documents as a Python snippet: AutoProcessor with use_fast=False, AutoModelForCausalLM in bfloat16 with device_map="auto", trust_remote_code=True, and audio passed through processor(messages) with audio_feats cast to the model dtype.

So the honest status is: the model loads, the serving scripts do not currently exist for a reader to read, and the card is describing a repository that has not been published yet. That is not a scandal — documentation landing a few hours before a code push is ordinary — but it does mean the two deployment recipes below cannot be audited, only quoted. Anyone weighing this checkpoint should plan on the Transformers path, and should read modeling_bailing_moe_v3_audio.py before enabling remote code, which the card itself advises.

What can be checked is the tokenisation contract, and the card is unusually specific about it: the validated ASR prompt asks the model to name the spoken language and put a tab between the language label and the transcription. If you are writing an evaluation harness, that tab is the parser. A generic "transcribe this" prompt is not the configuration the vendor measured.

The table inclusionAI published, including the columns it loses

Three PNG images in assets/ carry the evaluation. Every number on them is inclusionAI's own run; the asterisks mark entries the vendor labels self-evaluated, and none of it has been reproduced by a third party. Artificial Analysis publishes audio leaderboards and has no entry for this model. There is no arena score, no independent Elo, and no second opinion — so read the following as a vendor describing its own model, which is exactly what it is.

The speech-recognition image is the strongest part of the case, and it is strong on the axis the model was clearly built for. Averaged over its Chinese test set the flash model reports 3.45 against 3.54 for Qwen3.8-Omni Flash and 9.83 for Gemini 3.8 Flash; over English, 3.69 against 4.02 and 5.05. On dialects the gap widens sharply:

• Dialect average — Ming-3.0-audio-flash 9.40, Seed2.0 Lite 13.01, Qwen3.8-Omni Flash 11.58, Gemini 3.8 Flash 48.41.

• Hokkien — 12.84 against 49.95, 28.08 and 81.62 for the same three.

• Shanghainese — 11.15 against 26.32, 21.10 and 88.64.

• Cantonese — 4.09 against 6.64, 4.76 and 42.62. This one is close to a tie with Qwen3.8-Omni Flash and a rout of Gemini 3.8 Flash.

• Hubei — 9.46 against 11.35, 7.75 and 48.24. Qwen3.8-Omni Flash wins this row, and the vendor printed it anyway.

• Hunan — 8.76 against 20.30, 17.11 and 71.39.

The same image is less flattering where the task stops being transcription. On speaker diarization, AISHELL-4 reports a diarization error rate of 8.58 for the flash model against 2.80 for Qwen3.8-Omni Flash; AliMeeting, 8.66 against 3.40; AMI, 25.29 against 11.16 for MiMo-V2.6 Pro. If your problem is "who spoke when" rather than "what was said," this is not the model the vendor's own table recommends. The long-form concatenated minimum-permutation character error rate on AMI tells the same story: 39.62 for the flash model against 30.91 for MiMo-V2.6 Pro and 23.19 for Qwen3.8-Omni Flash.

The audio-intelligence image is where the case thins. The model wins its headline rows — BigBenchAudio 99.40, MMAU 82.00 — and then gives ground almost everywhere the task requires reasoning across a long recording rather than describing a short one.

• LongAudioSpan — 41.12, against 50.26 for Seed2.0 Lite and 66.40 for Qwen3.8-Omni Flash.

• Long-audio Chain score — 12.31, against 34.54 for Qwen3.8-Omni Flash, 48.70 for its thinking mode, and 49.63 for Gemini 3.8 Flash. That is a fourfold gap on the metric the vendor chose to publish.

• MMAR — 74.40 against 77.70 for Seed2.0 Lite. MMSU — 78.82 against 82.10. AudioMarathon — 68.41 against 70.06.

• Tool calling — BFCL-v3 Audio 81.40 against 75.08 for Seed2.0 Lite and 86.68 for Gemini 3.8 Flash.

The text benchmarks are the least surprising and the most useful, because they say what the model is not. On general reasoning the flash model lands below the omni models it is being compared with on nearly every row: GPQA Diamond 82.45 against 91.00 for Qwen3.8-Omni Flash, MMLU-Pro 80.00 against 87.70, IFBench 59.00 against 81.50, LiveCodeBench v6 79.90 against 92.60, AA-LCR 73.90 against 80.00. The backbone is a 6.5B-active audio specialist wearing a 126B parameter count, and the text table is what that looks like from the outside. If you want one model that both hears a meeting and writes the follow-up code, inclusionAI's own numbers say you should not expect this one to carry the second half.

What it costs to run

Two serving profiles are quoted, neither auditable yet. The reference vLLM-Omni profile for 128K context uses tensor parallelism across eight 80 GB H800 GPUs. The validated SGLang profile uses four GPUs and returns text only. Those are the vendor's configurations, not measurements, and with 256 GB of weights before any KV cache there is no smaller version of this — the 8K config the repository shipped in September would not have changed the memory floor, because the floor is the checkpoint.

The practical read is that this is a cluster model, not a workstation model. Four GPUs is roughly the cost of a serious GPU node; eight at 80 GB is a different budget line entirely. Ming-3.0-audio-tiny is the answer for anyone who wants to test the family without that commitment — 21 GB and 2.7B active parameters will fit on a single large card — but the tiny model is a different checkpoint with its own, consistently worse numbers on the same tables (long-audio Chain score 10.56 against 12.31, LongAudioSpan 35.45 against 41.12, URO-Bench-Pro 54.15 against 81.92, VoiceBench 74.29 against 86.74), and the vendor publishes nothing establishing that its behaviour transfers.

Worth noting for anyone planning an integration: the Talker is optional and separable. The card describes it as generating speech from the model's output, and the presets sit in their own folder with their own requirements file, so a text-only deployment does not have to carry it. Both documented serving profiles return text by default.

Where a router earns its place, and where it does not

OrcaRouter does not serve Ming-3.0-audio-flash. It is not in our catalogue, we cannot give you an endpoint for it, and nothing in this piece should be read as an offer to. What a router is good for in a pipeline built around a model like this is the half that is not the model. Speech systems are mostly text work: the prompt that decides what to ask the audio, the post-processing that turns a tab-delimited language label into structured output, the summary written from the transcript, the fallback that runs when the GPU node is saturated. Those are ordinary text calls, and one API for more than 200 models at 0% markup — provider list price passed through, so a vendor price cut is live on our side the same day — is a simpler thing to wire than four vendor contracts. Automatic failover is what lets you put a new, unproven model behind part of your traffic without betting the whole path on it.

The boundary is worth stating plainly, because it is the whole reason this article does not end with a sign-up box: a 259 GB checkpoint you run on your own eight GPUs is infrastructure, not an API call. A router routes calls. It does not make a cluster model into a hosted one, and it does not turn 22 downloads into production evidence.

What to watch

Three things would change the picture, and each is checkable in a way the benchmarks are not. First, the GitHub repository: when inclusionAI/Ming-3.0-audio-flash stops returning 404, the serving scripts and the voice guide become real and the two deployment recipes above become auditable. Second, an independent score: the moment Artificial Analysis or an arena lists this checkpoint, the fourfold long-audio gap stops being a vendor's self-report and starts being a data point. Third, the licence file: MIT is declared in metadata, and a LICENSE file in the repository would make it a fact rather than a claim — the distinction matters to anyone whose procurement process reads the repo rather than the card.

Until then the accurate description is the unglamorous one. A serious audio-language checkpoint, strongest on accented and dialectal Chinese speech, weak on long-audio reasoning and speaker diarization by its own published numbers, MIT-licensed on paper, roughly 259 GB on disk, and documented only as of this week. Six weeks of silence is not evidence of anything except that inclusionAI shipped the weights before it shipped the paperwork — but the paperwork is the part that tells you what you are downloading, and it is now, finally, there to read.