A generated title card reading 'Ming-3.0-audio-flash tops speech reasoning at 99.8%', with a subtitle noting the first-place rank on Artificial Analysis Big Bench Audio on 18 November 2026 and that the checkpoint was released on 3 September with no rate card. Four stat chips read '99.8% Big Bench Audio', '#1 of 42 models', '1.12 s to first audio' and 'no price published'. The footer reads 'Scores per Artificial Analysis, 18 November 2026.'
Engineering & Research

Ming-3.0-audio-flash Tops Artificial Analysis Speech Reasoning at 99.8% — One Point Ahead, and Still No Price

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Artificial Analysis now lists Ming-3.0-audio-flash first on its Speech Reasoning board, the Big Bench Audio evaluation, at 99.8% accuracy, ahead of StepAudio 3 Realtime at 99.7%, Qwen Audio 3.0 Realtime Plus at 99.2%, Eleven Realtime v1 Max at 98.8%, Grok Voice 3.0 at 98.5%, Gemini 3.8 Live Extended Thinking (High) at 97.7% and GPT-Realtime-2.1 High at 95.8%. The claim went out from Ant Ling on 18 November 2026 at 08:37 UTC, nineteen minutes after the model's own repository was edited to add the same figure and a leaderboard image. That timing is the whole story: a model that shipped in September with no third-party score now has one, and the number is very close to the ceiling.

To be precise about what is new and what is not — because this is not a launch. inclusionAI uploaded the Ming-3.0-audio-flash weights to Hugging Face on 3 September 2026 and said nothing about it for six weeks; the model card, the serving profiles and the example code only landed between 15 and 19 October. What changed today is the evaluation, not the checkpoint. The safetensors shards are the same bytes they were in September. So the honest framing is that a seven-week-old open-weights audio model has just been measured by someone other than its author, and it came out on top.

What Big Bench Audio actually asks a model to do

It is worth being concrete about the test, because "speech reasoning" sounds like something a voice agent does in conversation and mostly is not. Big Bench Audio is a set of multiple-choice questions delivered as audio: the model hears a recording, hears or reads a question about it, and picks an answer. It rewards the ability to hold a spoken passage in context and apply ordinary reasoning to it. It does not reward turn-taking, interruption handling, backchannels or the ability to sound like a person. Those live on a different Artificial Analysis board, Full Duplex Bench, where Ming-3.0-audio-flash currently has no published score at all.

That distinction matters more than usual here, because at 99.8% the metric is nearly out of room. On a 42-model board, the difference between first and fifth place is one percentage point. Between first and the model in thirteenth place — GPT-Realtime-2.1 High — it is four points. A test that almost everything modern scores in the nineties on is a test that has stopped separating the top of the field, and a one-point lead over StepAudio 3 Realtime is a handful of items on a benchmark whose item count nobody outside Artificial Analysis can check. Read the ranking as "this model belongs in the top group", which is genuinely useful, rather than "this model is measurably better at hearing than everything else", which the number cannot support.

The board in order, for anyone who wants the shape of the field rather than the headline:

• 1 — Ming-3.0-audio-flash, 99.8%

• 2 — StepAudio 3 Realtime, 99.7%

• 3 — Qwen Audio 3.0 Realtime Plus, 99.2%

• 4 — Phonic 1 (High), 98.9%

• 5 — Eleven Realtime v1 Max, 98.8%

• 7 — Grok Voice 3.0, 98.5%

• 8 — Gemini 3.8 Live Extended Thinking (High), 97.7%

• 13 — GPT-Realtime-2.1 High, 95.8%

A screenshot of the Artificial Analysis Speech to Speech page, showing the site navigation, the page heading, a blurb describing comparison of speech-to-speech models across reasoning quality, conversational dynamics, generation time and price, and two chart panels labelled Speech to Speech Index and Speed & Cost per Hour of Input Audio, plus a Related Links rail. The per-model rows inside the charts are not legible at this size.

One detail in the vendor's own paperwork is worth flagging, because it is the kind of thing that gets quietly dropped. The evaluation images on inclusionAI's model card, uploaded in October, show the flash model at 99.40 on "BigBenchAudio". Artificial Analysis now reports 99.8. Those are two different runs of a similar benchmark, and the gap is not an error in either — it is the normal spread you get when the harness, the prompt format and the item subset are not identical. It is also a reminder that the vendor's October table and today's third-party board are not interchangeable numbers, even when they carry almost the same label.

What the board does not say

Three things are missing from the Artificial Analysis entry, and each of them matters more to a deployment decision than the accuracy figure.

The first is price. Ming-3.0-audio-flash has no per-token or per-minute rate on Artificial Analysis because there is nothing to buy: it is a 126-billion-parameter mixture-of-experts checkpoint under MIT, with 6.5 billion parameters active per step. The cost is whatever the hardware costs. inclusionAI's own deployment notes put the reference 128K serving profile at tensor parallelism across four 80 GB H800s, with an eight-GPU fallback for long audio, and a validated SGLang profile at four GPUs returning text only. Every model above it on the leaderboard except itself is a metered API. Comparing them on accuracy alone compares a benchmark score to a benchmark score and ignores that one side of the table is a purchase and the other is a subscription.

The second is latency. Artificial Analysis does measure time to first audio, and Ming-3.0-audio-flash sits at 1.12 seconds. Eleven Realtime v1 Max is at 0.57 seconds — roughly half. If your application is a voice agent that has to answer inside a conversational pause, a benchmark win does not compensate for a second of dead air, and the leaderboard's ordering says nothing about which of those two numbers you can live with.

The third is conversational behaviour. Full Duplex Bench scores exist for Eleven Realtime v1 Max (0.941 overall, with interruption handling at 0.99 and turn-taking at 1.0) and for several Gemini and GPT realtime models. Ming-3.0-audio-flash has no FDB score, no conversational-dynamics breakdown, and no published measurement of how it handles being interrupted. A model can ace multiple-choice questions about audio and still talk over you.

And the vendor's own October numbers are the least flattering evidence in this whole picture. On the same card that advertises the speech-reasoning win, inclusionAI published a long-audio chain score of 12.31 against 34.54 for Qwen3.8-Omni Flash and 49.63 for Gemini 3.8 Flash, a LongAudioSpan of 41.12 against 66.40, and speaker-diarization error rates on AMI of 25.29 against 11.16 for MiMo-V2.6 Pro. Those are the vendor's figures, on the vendor's chosen tasks, and they describe a model that is excellent at short, question-and-answer audio and weak at reasoning across a long recording. Nothing about a 99.8 on Big Bench Audio changes that, because Big Bench Audio is not a long-audio test.

The seventeen-minute gap between card and tweet

The repository commit is timestamped 08:26:48 UTC on 18 November and is titled, in part, "Update Flash BBA score to 99.8% and add the overall #1 leaderboard res…". The post from Ant Ling followed at 08:37:02. The same commit touched the tiny sibling's card forty-eight seconds later, adding the same sentence — a shared benchmark paragraph, copied to both repositories, and one more reason to check which checkpoint a number actually belongs to before quoting it.

A generated single-column scoreboard titled 'Ming-3.0-audio-flash - the scoreboard'. Rows read 'Big Bench Audio: 99.8%, rank 1 of 42', 'Vendor card, October: 99.40, self-evaluated', 'Time to first audio: 1.12 s', 'Full Duplex Bench: not published', 'Price: no rate card' and 'Reference hardware: 4 x 80GB H800'. The footer reads 'Big Bench Audio score, rank and latency per Artificial Analysis, 18 November 2026; the 99.40 figure is vendor-reported and unaudited.'

What that sequence tells you is that the model card is now downstream of the leaderboard rather than the other way round. For six weeks the only performance evidence for Ming-3.0-audio-flash was a set of PNG images in an assets/ folder, produced by the team that trained it. That is not nothing — vendor evaluations are usually honest, just optimistic about task selection — but it is a different category of evidence from a third party running its own harness and publishing a rank. The interesting thing about today is not that the vendor claims first place. It is that an organisation with no stake in the result put the model there.

A screenshot of the inclusionAI Ming-3.0-audio-flash model card on Hugging Face, showing the audio-to-audio pipeline tag, the MIT licence badge, a reported model size of 128B params, a downloads-last-month figure of 2,572, the headline 'Hear clearly. Understand deeply. Act intelligently.', and a model summary table listing the developer as inclusionAI, the model type as a native audio-language mixture-of-experts, and 126B total parameters with 6.5B activated per inference step.

Where a router fits, 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 here should be read as an offer. If you want to run it, you clone the repository, install the Transformers 5 code path or one of the SGLang-Omni serving profiles, and you point at least four 80 GB cards at it. That is infrastructure, and a router does not do infrastructure.

What a router does do is the other half of any speech pipeline. A voice application built on a self-hosted audio model is still mostly text work: the prompt that decides what to ask the recording, the post-processing that turns a tab-delimited language label into structured output, the summary written from the transcript, the classifier that routes a noisy call to a human. Those are ordinary chat 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 a second and third vendor contract. Automatic failover is what lets you put an unproven open-weights model behind part of your traffic without betting the whole path on it.

The boundary is the point. A 259 GB checkpoint you run on your own GPUs is a capital decision; a router routes calls. It does not turn a cluster model into a hosted one, and it does not make 99.8% on a multiple-choice benchmark into evidence about your product.

What would make this mean something

Three measurements would turn today's headline into a decision, and all three are things Artificial Analysis or an arena could publish without asking inclusionAI for anything. A Full Duplex Bench score would say whether the conversational behaviour matches the comprehension. A long-audio result on a public set would test the one axis where the vendor's own table is weakest. And a second independent run of Big Bench Audio — a different harness, a different item sample — would tell you whether 99.8 is a ceiling effect or a real gap.

Until then the accurate sentence is narrow and worth saying plainly: an open-weights audio model from Ant Group, released in September and documented in October, is now ranked first on a third-party speech-reasoning benchmark by a margin of one percentage point over the next model, with no published price, twice the time to first audio of the fastest competitor on that board, and no conversational score at all. That is a real result. It is also a much smaller claim than "best voice model", and the difference between those two sentences is where most of the useful thinking about this model still has to happen.

Compared in this article2

Detected from this article · Benchmarks: Artificial Analysis · updated daily