Ming-3.0-asr-tiny vs Granite Speech 5.0 470M TurboCTC: 19× the Parameters, None of the Throughput

Put Ming-3.0-asr-tiny and Granite Speech 5.0 470M TurboCTC on the same page and the first number that jumps out is the parameter ratio: 8.87 billion against 470 million, roughly nineteen to one. The second is that the small model is the one built for speed by an enormous margin, and the third is that this is not a contradiction. Granite Speech 5.0 470M TurboCTC has no decoder to run — it is an encoder that emits characters and stops. Ming-3.0-asr-tiny has a 24-layer mixture-of-experts decoder that writes one token at a time and can be told, in advance, which names are about to be said. One is a throughput machine for English audio. The other is a context machine for Chinese audio. IBM released Granite Speech 5.0 470M TurboCTC on 25 August 2026 with a blog post, a technical report and an Open ASR leaderboard entry; inclusionAI put Ming-3.0-asr-tiny's weights on Hugging Face on 14 October 2026 with no announcement at all. Every Granite figure used here comes from the Granite Speech 5.0 470M TurboCTC model card, which carries the Open ASR table, the noise sweep and the single-language statement. 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.

Neither model is in the OrcaRouter catalogue, and neither appears in the other's evaluation, so there is no vendor head-to-head to quote. Everything below is drawn from each model's own published configuration and charts, labelled as such.

The architectural fork, stated plainly

Granite Speech 5.0 470M TurboCTC is an encoder-only model, and IBM says so as a deliberate break from its own past. Previous Granite Speech models were an acoustic encoder, a projector and a Granite language model with LoRA adapters — which is to say, they were language models wearing a microphone. The 5.0 generation drops the language model entirely and optimises a CTC objective over a 16,384-unit BPE head. Sixteen conformer blocks, 1,024 hidden dimensions, eight attention heads, block attention over 128 frames, self-conditioning at the eighth block, and a 7-wide convolution kernel.

The subsampling is the part that produces the speed. Audio arrives as log-mel plus deltas at 100 frames per second in 320 dimensions. Three stages of 2× reduction — first a reshape that stacks consecutive frames, then strided convolutions inside the first two conformer blocks — bring that to 12.5 output tokens per second. IBM's blog is explicit that this is the change from its earlier encoders, which emitted 50 characters per second: four times fewer output steps per second of audio, and no autoregressive loop deciding each one.

Ming-3.0-asr-tiny went the other way on every axis that matters here. Its decoder is 24 layers of mixture-of-experts with 128 experts, eight routed per token, a shared expert and a dense first layer, alternating three K​imi Delta Attention layers with one Gated MLA layer. It writes tokens autoregressively. It also carries a 32-layer, 1,280-dimension Whisper-shaped encoder — the shape of Whisper large-v3, with the repository shipping modeling_whisper_encoder.py and O​penAI's MIT notice in a separate file. Ming kept the language model and made it the expensive part.

Where the numbers actually come from

This is the section that decides how much weight either model's claims deserve, and the two releases are not comparable in kind.

• Granite Speech 5.0 470M TurboCTC — 5.00% aggregate WER across the public English short-form sets on the Open ASR leaderboard, as of 25 August 2026, with 1.93% on LibriSpeech test-clean and 3.59% on test-other. Its non-commercial sibling scores 4.85%. Throughput is over 12,600 RTFx on a single H200 with batched inference. There is a technical report on arXiv and a Hugging Face blog post, and the results are submitted to a public leaderboard that lists them.

• Ming-3.0-asr-tiny — 3.55 average CER on Chinese public sets, 3.99 average WER on English public sets, 8.42 on fifteen dialect sets, 1.67 across five in-house domains, 4.36 on a context-ASR aggregate. All of it exists as PNG images on a model card. There is no technical report, no leaderboard submission, and no evaluation script in the repository.

That blog post is worth opening for the same reason the model card is: it is the trail Ming does not have. IBM published a dated article, a technical report and a leaderboard entry, and the throughput claim in the article can be checked against the batch configuration it describes.

That asymmetry is not a knock on Ming's numbers; it is a statement about what you can check. Anyone can re-run Granite's Open ASR figures from public data in an afternoon. Re-running Ming's requires reconstructing its test sets, including five in-house domains whose composition is not published.

One more asymmetry, and it runs the other way: the two models do not share a single evaluation set. Granite's charts are English-only, spanning LibriSpeech, AMI, Earnings-22, GigaSpeech, SPGISpeech, TED-LIUM and VoxPopuli. Ming's English average uses LibriSpeech, Multilingual LibriSpeech, FLEURS English, Common Voice 15 English and VoxPopuli — overlapping on two of five, with different aggregation. A reader who wants "which is more accurate on English" will not get an answer from either vendor, because neither asked the question.

Speed means two different things here

IBM's 12,600 RTFx is a batched throughput number: audio pushed through a GPU in bulk, more than three and a half hours of speech per second of compute. It is the right number for transcribing an archive, and it is close to meaningless for a live microphone. IBM does ship streaming — a WebGPU demo in the browser, chunkwise attention, and an MLX path for Apple Silicon — but the headline figure describes the batch case.

Ming publishes no RTFx at all. What it publishes is a size: 17.7 GB of BF16 weights and an autoregressive decoder with 128 experts. The token-by-token loop is the cost model. Every extra second of audio means more decoder steps, and every decoder step touches a routing decision across 128 experts. That is the structural reason a 19× smaller encoder-only model is the one with a throughput claim, and it does not need a benchmark to be true.

The practical consequence is that these are not competing for the same deployment. Granite Speech 5.0 470M TurboCTC is designed to sit on a laptop, a phone, or a single H200 chewing through a podcast archive. Ming-3.0-asr-tiny 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, and a separate SGLang adapter pinned to a specific commit if you want to serve it. One of those is a dependency list you can put in a product.

The capability Granite gave up to get there

The encoder-only design buys speed and size, and IBM is candid about the bill. Its blog states that the 5.0 models give up capabilities the LM-equipped versions had, naming speech translation and keyword biasing. That second one is the interesting one, because keyword biasing is precisely what Ming calls Context ASR and treats as its headline feature.

Ming accepts a list of hotwords or the preceding turn of dialogue as part of the transcription prompt, and the vendor's own example is instructive: an audio clip where the film title 同囚 comes out as 同球 on the plain path and correctly once the topic and a hotword list are supplied. The card is careful to say hotwords are hints rather than guarantees, which is the honest framing — biasing shifts probability, it does not enforce output.

Granite Speech 5.0 470M TurboCTC has no equivalent. There is no prompt, no context field, no list of names to supply, because there is no language model to condition. If your audio is dense with product names, drug names, or surnames from a customer database, the model has exactly one tool: whatever it learned in training. IBM's answer to that class of problem is a notebook for fine-tuning on your own data, which is a different workflow with a different cost.

Language coverage runs the same way. Granite is English, stated as the only supported language, trained on roughly 60,000 hours of English audio in ten days on eight H100s. Ming is Chinese and English with fifteen regional Chinese dialect test sets including Cantonese, Minnan and Shanghainese, and a dialect average of 8.42 that its own charts place ahead of every open baseline it lists.

Choosing between them

If your audio is English and your problem is volume — call recordings, archives, meeting transcripts, anything where cost per hour outweighs the last point of accuracy — Granite Speech 5.0 470M TurboCTC is the more sensible starting point, and the fact that it runs quantised through a native C++ runtime without Python is a deployment advantage no amount of Chinese dialect coverage offsets.

If your audio is Mandarin or a regional dialect, or if the accuracy problem is proper nouns you can list in advance, Ming-3.0-asr-tiny is the only one of the two that addresses it, and it does so with an architecture built for it rather than a fine-tune bolted on. It is also five days old with 35 downloads, which means you would be among the first to discover how it fails.

There is a third case worth naming: you want both, and you want to know which is winning on your own traffic. That is the routing problem OrcaRouter is built for. Neither of these models is hosted there — Ming-3.0-asr-tiny and Granite Speech 5.0 470M TurboCTC are both self-hosted open weights — but the same endpoint that fronts your hosted models can front a candidate ASR path, with automatic failover so a failed request falls back to the incumbent instead of surfacing to a user, and one key across 200-plus models so the comparison does not require a second contract.

What neither vendor has shown you

Neither model has been scored by anyone independent. Granite's figures at least sit on a public leaderboard that other teams also submit to, which means they can be argued with; Ming's sit in four PNGs that were still being edited on the afternoon they were published. Neither card documents word-level timestamps or speaker diarisation, so any pipeline that needs timing is out of scope for both — Granite's card describes plain transcription output, and Ming's usage notes say explicitly that output is transcription text without word-level timestamps or diarisation.

The comparison that would actually settle this cannot be run by either vendor. It needs one corpus, one language pair, one scoring script, and both models on the same GPU. Until somebody publishes that, the only defensible statement is the one the architectures already make: a nineteen-times-larger autoregressive mixture-of-experts model is not competing with an encoder-only CTC model on throughput, and an encoder-only CTC model cannot be told what words to expect.