Ming-3.0-asr-tiny vs NVIDIA Parakeet TDT 0.6B: A 35-Download Challenger Meets a 626,000-Download Incumbent

The download counters are the first thing to look at, because they describe a gap that no benchmark in this article can close. NVIDIA Parakeet TDT 0.6B v3 sits on more than 626,000 downloads and 1,200 likes on Hugging Face, released on 14 August 2025 and shipped inside the NeMo toolkit, a native C++ runtime and a Transformers path people actually use. Ming-3.0-asr-tiny sits on 35 downloads and a handful of likes, with weights that appeared on 14 October 2026 and no announcement from its vendor. One is infrastructure. The other is a checkpoint five days old. But the two are worth comparing, because they solve different problems and the interesting part is that Ming's problem — Chinese dialects and hotwords — is one Parakeet TDT 0.6B v3 does not attempt. Every Parakeet figure here, including the 626,738 download count, is from the NVIDIA Parakeet TDT 0.6B v3 model card, which also carries the Open ASR table and the MUSAN noise sweep. 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.

NVIDIA's card is unusually complete by the standards of this blog's usual subjects: a full Open ASR table, a noise-robustness sweep, a per-language breakdown across 25 European languages, and a stated release date. Ming's card is four PNG charts and no technical report. That difference in documentation is itself a finding, and it belongs in the decision alongside the accuracy numbers.

Two decoders that are both not-a-language-model

Parakeet TDT 0.6B v3 is a FastConformer encoder with a Token-and-Duration Transducer decoder — 600 million parameters, 16 kHz mono input, punctuation and capitalisation in the output. TDT is the reason it is fast: instead of emitting one token per frame, the transducer predicts both a token and a duration, so the model can skip frames and advance several at a time. It is not an encoder-only CTC model like IBM's Granite Speech line, and it is not a language model with an audio front end like Ming. It is the middle case, and it has been the middle case long enough to be boring, which is a compliment.

Ming-3.0-asr-tiny is the opposite bet. Its decoder is 24 layers of mixture-of-experts — 128 experts, eight routed per token, one shared, the first layer dense — alternating three K​imi Delta Attention layers with one Gated MLA layer. Its encoder is 32 layers at 1,280 dimensions with 20 heads and 128 mel bins, which is the Whisper large-v3 shape, and the repository ships modeling_whisper_encoder.py with O​penAI's MIT notice to make the lineage explicit. The model is 8.87 billion parameters by its safetensors index, about 2 billion of them active per token.

That makes the parameter comparison lopsided in a way that does not translate into accuracy: 8.87 billion total against 600 million, and roughly 2 billion active against a dense 600 million. Ming is a language model wearing a microphone. Parakeet is a speech model that learned to read.

The things Parakeet does that Ming does not

Three capabilities separate them before accuracy is discussed, and all three are documented on the cards rather than inferred.

The first is timestamps. Parakeet TDT 0.6B v3 returns character-level, word-level and segment-level timing through both its NeMo and Transformers paths, and the card shows the call that produces them. That single capability is what subtitle generation, forced alignment, meeting diarisation and any "jump to the moment this was said" interface are built on. Ming's usage notes state that output is transcription text without word-level timestamps or speaker diarisation. This is not a gap in a benchmark; it is the absence of an output field.

The second is long audio. Parakeet handles it through local attention — the card's own recipe swaps in relative-position local attention with a context of 256 frames either side for long-form transcription, and NVIDIA documents limits of roughly 24 minutes with full attention on an 80 GB A100 and three hours with local attention. Ming accepts an 8,192-position context and chunks longer audio through a command-line argument, which is a workable but different kind of answer.

The third is the ecosystem. Parakeet runs in NeMo, in a Transformers path with AutoModelForTDT, and in a quantised GGUF through NeMo-Speech.cpp with no Python at runtime. Ming needs Python 3.12, PyTorch 2.11.0, torchaudio 2.11.0, Transformers 5.12.1, fla-core 0.5.2, CUDA, and trust_remote_code=True, plus a separately pinned SGLang adapter if you want to serve it. On a machine with no GPU, that difference is decisive before any WER number is read.

Read that list against Ming's card and the comparison stops being about accuracy. Every line in it is a capability Ming's repository explicitly says it does not have, stated by NVIDIA as a feature and by inclusionAI as a limitation.

What each vendor measured, and on what

Parakeet's English numbers come from the Open ASR leaderboard public sets: 6.34% average WER, with 1.93% on LibriSpeech test-clean, 3.59% on test-other, 11.31% on AMI, 11.42% on Earnings-22, 9.59% on GigaSpeech, 3.97% on SPGI Speech, 2.75% on TED-LIUM and 6.14% on VoxPopuli.

Its multilingual story is thinner than the language list suggests, and this is worth saying plainly because the card does not. The 25 languages are real, but the published multilingual evaluation is FLEURS, MLS and CoVoST averages — 11.97%, 7.83% and 11.98% — and the Open ASR table above is English-only. A team evaluating this for Polish or Finnish gets a per-language FLEURS number and no long-form or noisy evaluation in that language at all.

Noise robustness is where the card earns its keep. Across MUSAN noise at descending signal-to-noise ratios, the average WER goes 6.34 clean, 7.12 at 10 dB, 8.23 at 5 dB, 11.66 at 0 dB and 19.88 at −5 dB — a relative degradation NVIDIA states as −213.64% at the worst point. That is a real curve, on public data, and it tells you something a single accuracy figure cannot: below about 5 dB this model degrades faster than linearly.

Ming's numbers, by contrast, are a Chinese average CER of 3.55, an English average WER of 3.99, a fifteen-dialect average of 8.42, a five-domain in-house average of 1.67 and a context-ASR aggregate of 4.36. All vendor-reported, all unreproduced, and none on a shared test set with Parakeet. On the one set both report — LibriSpeech test-clean — Ming does not publish a per-set figure in prose, only its five-set English average, so the comparison cannot be made from the cards at all.

Licences are not equivalent here

Ming's weights and original code are MIT, with bundled Apache-2.0 files retaining their own headers and the Whisper-adapted encoder keeping O​penAI's MIT notice. Parakeet TDT 0.6B v3 is CC BY 4.0, which permits commercial use with attribution. Both are permissive; neither is a trap. The practical difference is attribution language in your product documentation, not legal exposure.

What is worth flagging is that the licence on Ming's repository is more permissive than its distribution story. MIT weights that require trust_remote_code=True and a pinned dependency set are permissive to use and awkward to deploy, and the two facts are independent.

The context feature is the only clean differentiator

Parakeet TDT 0.6B v3 has no prompt, no hotword list and no way to supply the names that will appear in the audio. It transcribes what it hears using what it learned across 25 European languages, and for domain vocabulary the documented route is fine-tuning, not conditioning. NVIDIA also ships a chunked-streaming script with configurable left and right context, which is a latency tool rather than an accuracy one.

Ming accepts hotwords or the preceding dialogue turn as part of the transcription prompt, and its dialect coverage — fifteen Chinese test sets including Cantonese, Minnan and Shanghainese, averaging 8.42 — is a capability Parakeet does not claim at all. If your audio is Mandarin, or a regional variety of it, or a Chinese business call full of product names, Parakeet is not a candidate and Ming is the only one of the two on the table. That is a real, specific, checkable advantage, and it is the reason this comparison exists rather than being settled by the download counter.

The caveat is that hotwords are hints. Ming's card says so directly: supplying a term raises the chance of it being produced and does not guarantee it. Anyone who has used keyword biasing on any ASR system knows the failure mode — the model starts hearing the hotword in audio that does not contain it. Test that before trusting it.

What to do with this

If you need English transcription with timing, on hardware you already own, with a runtime that does not require a Python environment, Parakeet TDT 0.6B v3 is the safe answer and the download count is not vanity — it is a proxy for how many failure modes have already been found and written down by other people.

If you need Chinese, dialects, or hotword conditioning, Ming-3.0-asr-tiny is the only one of the two that can do the job, it is genuinely novel in that niche, and it is five days old with essentially no external testing. The right posture is a bake-off on your own corpus, with the timestamp gap treated as a hard constraint rather than a trade-off: if your downstream system consumes word timings, the evaluation is over before it starts.

For teams running that bake-off across a production estate, the comparison usually ends up inside a router rather than in a spreadsheet — one endpoint in front of self-hosted candidates, automatic failover so a failed request falls back to the incumbent, and one key across 200-plus hosted models instead of a second contract for every comparison. Neither Ming-3.0-asr-tiny nor NVIDIA Parakeet TDT 0.6B v3 is in the OrcaRouter catalogue, so neither is something we serve; the routing layer earns its place around them, not through them.

The asymmetry that will not go away

One model has a release date, an arXiv-style documentation trail, a leaderboard submission, a noise sweep and a per-language table. The other has four charts, a config file, a commit log and a release film. That is not a criticism of Ming's engineering — the architecture is more ambitious than Parakeet's and the dialect bet is real — but documentation is what turns a checkpoint into something a team can adopt without becoming the vendor's QA department.

The most useful thing either vendor could publish next is the same thing: a shared evaluation on Chinese and English audio, scored by a third party, with timestamps measured as well as words. Until one exists, the decision rests on capability boundaries that are already documented — timing, streaming, hardware, languages, hotwords — and those boundaries are sharp enough to choose on.