Ming-3.0-asr-tiny vs Whisper large-v3-turbo: Whisper's Encoder, a Different Mouth

Open the config file for Ming-3.0-asr-tiny and the audio tower reads back at you: 1,280 hidden dimensions, 32 encoder layers, 20 attention heads, 128 mel bins. Those are the dimensions of Whisper large-v3's encoder, and the repository says so out loud — it ships a file called modeling_whisper_encoder.py and keeps O​penAI's MIT notice in a separate WHISPER_LICENSE. This is therefore not a story about two unrelated speech stacks. Ming-3.0-asr-tiny is a Chinese-first recogniser that kept Whisper's ears and replaced the mouth: a 24-layer mixture-of-experts decoder, 8.87 billion parameters according to its safetensors index, conditioned on hotwords and preceding dialogue. Whisper large-v3-turbo is the 809-million-parameter distillation that kept the same encoder and cut the decoder from 32 layers to four. Every Ming figure in this article comes from one place, and it is the Ming-3.0-asr-tiny model card on Hugging Face, which holds the weights, the config, the four evaluation charts and a release film asset dated today. The 809M parameter, 99-language, MIT-licensed baseline is documented at the Whisper large-v3-turbo model card, together with the word-timestamp and translation behaviour this comparison keeps coming back to. 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.

If you want the short answer before the detail: they are not substitutes. Whisper large-v3-turbo emits timestamps, translates into English, covers 99 languages, and runs on hardware that is not a CUDA GPU. Ming-3.0-asr-tiny emits plain text and nothing else, covers Chinese and English with a serious bet on regional dialects, and needs a recent PyTorch and Transformers stack with remote code execution enabled. The overlap is Chinese transcription of short clips, and even there the two have never been scored against each other by anyone.

What shipped, and how quietly

The Hugging Face repository for Ming-3.0-asr-tiny was created on 14 October 2026. When this was written it showed 35 downloads and a single like. The companion GitHub repository has 20 commits, one star, no forks and an empty Releases tab. There is no vendor blog post, no press coverage, and no launch announcement of any kind — the model card is the announcement.

The card also declines to state a release date. The only dating signal in the repository is a release film asset filed under a path containing media-2026-10-19, which is today. Treat the weights as days old and the announcement as not yet made, rather than as a product launch you missed.

What the repository does establish, beyond doubt, is the shape of the thing you would be downloading: 8.87 billion parameters split across two safetensors shards in BF16, a 157,184-token vocabulary, a 24-layer text decoder, a 32-layer audio encoder, and an asr_output_format field set to plain. The repository's own benchmark charts label the model 8.6B-A2B, which is a vendor shorthand for roughly 2 billion parameters active per token; the safetensors index is the harder number, and the two differ by about 270 million.

Same ears, different mouth

The encoder is the part people skip past, and here it is the most informative line in the config. A 1,280-dimension, 32-layer, 20-head audio encoder with 128 mel inputs is the Whisper large architecture, and the file names in the repository confirm the lineage rather than merely resembling it. The encoder is also the smaller half of Ming: 32 layers of a conventional transformer against a decoder that carries 128 experts, routes eight of them per token, keeps the first layer dense, and compresses attention with low-rank projections (q_lora_rank 256, kv_lora_rank 512). That is a language model with an audio front end bolted on, not an ASR model with a language model attached.

Whisper large-v3-turbo went the other direction. It is the same encoder with a decoder distilled from 32 layers to four, which is where the 809 million parameter count and the roughly eight-times speed-up over large-v3 come from. Everything that makes turbo cheap sits on the decoding side; everything that makes Ming expensive sits there too.

• Parameters — Ming-3.0-asr-tiny 8.87B in BF16, of which the vendor brands about 2B as active per token. Whisper large-v3-turbo 809M.

• Decoder — 24 layers, 128 experts with 8 routed per token, low-rank compressed attention. Whisper: 4 layers, dense, distilled from large-v3's 32.

• Audio front end — a 32-layer, 1,280-dimension Whisper-shaped encoder in both cases, which is the least obvious fact in this comparison.

• Input window — Ming takes 3,000 mel frames up to 30 seconds and 6,000 beyond that, with longer audio chunked against an 8,192-position context. Whisper's receptive field is 30 seconds and chunking is handled by the surrounding tooling.

• Output — plain transcription text versus transcription plus sentence-level and word-level timestamps, plus translation into English.

• Languages — Chinese and English, with 15 regional Chinese dialect test sets, versus 99 languages.

• Licence — MIT for Ming's weights and original code, with bundled Apache-2.0 files keeping their own headers; MIT for Whisper large-v3-turbo. This is the one row where there is nothing to argue about.

What the vendor's charts show, and who is missing from them

Ming's evaluation lives entirely in four PNG charts on the model card. The repository text contains no numbers at all, and the charts repay reading closely because they are not uniformly flattering.

On the Chinese average — a CER mean over AISHELL-1, AISHELL-2, Common Voice 15 Chinese, FLEURS Chinese and WenetSpeech — Ming scores 3.55, and the best figure anywhere in that table is 2.94. On the English average — LibriSpeech clean and other, Multilingual LibriSpeech, FLEURS English, Common Voice 15 English and VoxPopuli — Ming scores 3.99 against a best-in-table 3.69. On the 15-dialect average Ming scores 8.42 against a best-in-table 8.00. In other words, on the vendor's own charts, Ming is consistently near the front of the field and consistently not first.

Where it looks strongest is the in-house domain sets, where CER runs 1.62 on chat, 1.03 on government, 1.58 on health, 2.49 on knowledge and 1.61 on local life. Those are the numbers to treat most carefully, because in-house test sets share an unseen amount of provenance with undisclosed training data, and the card says nothing about what the model was trained on.

The context-ASR table is the one that deserves a second look. Its aggregate puts Ming at 4.36, behind Qwen3-Omni at 4.34 and Qwen3-ASR-1.7B at 4.23, and ahead of Gemini 2.5 Pro at 6.3. The model's headline feature — feeding it hotwords and dialogue context — does not make it the best context-aware recogniser in its own comparison. It does win individual named-entity rows, particularly on entity false-negative rate, which is the metric that matters when the question is "did it get the name right" rather than "how many characters were wrong".

Now the omission that decides how much any of this is worth to you: Whisper large-v3-turbo does not appear in any of the four charts. The baselines are Q​wen omni models, Qwen3-ASR-1.7B, Fun-ASR 1.5 and FireRedASR2-LLM-8B. There is no vendor-reported head-to-head between the two models in this article, and any table that claims one has invented it. Every figure above is vendor-reported and has not been independently reproduced.

Where Whisper still cannot be followed, and where Ming cannot

Timestamps are the hard boundary. Whisper large-v3-turbo returns sentence-level and word-level timing alongside the text, which is what subtitle generation, forced alignment and diarisation pipelines are built on. Ming returns transcription text with no word-level timestamps and no speaker diarisation, and the card states this as a limitation rather than an omission. If your pipeline needs to know when a word was said, the comparison ends there — not because Ming scored worse, but because the output does not exist.

Streaming is the second boundary. The bundled example does not accept streaming input, and the model is designed around whole clips with an optional language hint. Whisper's ecosystem has spent two years building incremental and low-latency wrappers around a 30-second window; Ming has been public for days and has none.

Deployment is the third. Whisper large-v3-turbo at 809 million parameters fits comfortably on a single consumer GPU and has CPU inference paths through optimised runtimes. Ming needs a CUDA GPU, PyTorch 2.11.0, torchaudio 2.11.0, Transformers 5.12.1, fla-core 0.5.2, and trust_remote_code=True to load its own modelling file — which is a supply-chain decision as much as an engineering one. Roughly 17.7 GB of BF16 weights is not a laptop download.

Where Ming pulls ahead is everything the context mode exists for: Chinese speech with dialect coverage that Whisper handles poorly, and short utterances dense with names, titles and domain jargon that you can list in advance. A separate serving path exists through an SGLang adapter with a pinned commit, which suggests the vendor expects this to be served rather than only run as a script.

Applying this to a real decision

If your audio is English, or multilingual, or you need subtitles with timing, or you want something running on a CPU box tomorrow, the answer is Whisper large-v3-turbo and it is not close. Two years of production use, an ecosystem of runtimes, and an output format everything already consumes are worth more than a few CER points on a vendor chart.

If your audio is Mandarin or a regional dialect, if your clips are short, and if you can enumerate the names that will appear in them, Ming-3.0-asr-tiny is worth a bake-off — but only a bake-off. It has 35 downloads. Nobody outside the lab has published a failure mode for it, which means you would be the one discovering them.

Do not read the parameter count as a quality signal in either direction. Ming is eight times larger than turbo and does not beat it on any shared, independently measured axis, because no shared axis has been measured. The 8.87 billion parameters buy a language model's worth of contextual prior and a dialect bet, not a documented accuracy win.

Running the comparison on your own audio

Neither model is in the OrcaRouter catalogue today, so the honest version of this section is that you would get them from the vendor's own repository and run them yourself. What a routing layer is good for here is the part that comes after: putting a slice of your real traffic through a candidate model while the incumbent keeps serving the rest, with automatic failover if the new path fails, and one API key across 200-plus models so that adding a second opinion does not mean a second contract. OrcaRouter passes provider list prices through at 0% markup, which matters less for open weights you host yourself and more for the hosted models you would benchmark against them.

The test worth running takes an afternoon: 200 clips from your own corpus, half of them containing a proper noun, scored by hand. Give Ming the hotword list and give Whisper large-v3-turbo the same names in its prompt field. Compare named-entity recall rather than character error rate, because that is the axis Ming was actually built for and the one the vendor's context table reports. Then decide whether the timestamp gap is survivable — which, for most pipelines, it will not be.

What would change this verdict

Two things. An independent evaluation on a shared test set would settle the accuracy question that neither vendor has answered, and anyone can run it. And a release announcement — with a training-data statement, a stated release date, and a timestamp roadmap — would turn Ming from a checkpoint with a promising config into a product you could plan around. Until then the correct description is the unglamorous one: a well-documented set of weights, shipped quietly, with a Whisper encoder inside it and no one yet outside the lab who can tell you how it fails.