Ming-3.0-audio-flash vs InternLumina U2: Two Quiet Ships, Opposite Paperwork

Both of these models were published without a press release, and that is roughly where the similarity ends. Ming-3.0-audio-flash went up on Hugging Face on 3 September 2026 as 259 GB of weights with almost no documentation, and only received a real model card on 19 October. InternLumina-U2 went the other way round: Shanghai AI Laboratory published the inference code and a project page on 1 September 2026, promised the weights, and shipped them later in hardware-specific folders. One team pushed the artefact first and the explanation six weeks later. The other wrote the explanation first and made you wait for the artefact. Neither announced anything, neither has an independent benchmark, and neither is a substitute for the other — one hears, the other sees and draws. The documented AI failures and quiet-release incidents behind comparisons like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

That makes this a comparison about two things at once: two different bets on what an open multimodal model is for, and two different habits of releasing one. The second half is more useful than it sounds, because the paperwork is what tells you whether you can deploy the thing on a Tuesday. Everything below comes from the two repositories and their cards. Ming's numbers are inclusionAI's own; InternLumina's are Shanghai AI Lab's, and its own card labels them "preliminary, partial results." No third party has scored either.

The two release timelines, side by side

Dates are the most objective evidence available here, so start with them.

• 1 September 2026, 04:03 UTC — the GitHub repository InternLM/InternLumina-U2 is created. Thirteen minutes later a Hugging Face repository of the same name appears, holding a .gitattributes and a README stub. The release that day is code: one driver script with a section per task, a README in English and Chinese, and a roadmap.

• 3 September 2026, 11:45 and 11:49 UTC — inclusionAI creates Ming-3.0-audio-tiny and Ming-3.0-audio-flash on Hugging Face. Nine minutes later 25 shards land on the flash repository. The README that came with them was titled "Ming Audio V3" and described only the ASR path.

• 13 to 19 October 2026 — the Ming repositories move four times in seven days: a config that raises the context from 8,192 to 131,072, a re-upload that drops the video-language variant, a voice-preset folder, and finally the card, the assets and an MIT relicence. inclusionAI/Ming-3.0-audio-flash appears on GitHub on 15 October.

• Sometime after 24 September 2026 — internlm/InternLumina-U2 on Hugging Face stops being a stub and gains two checkpoint trees, ascend/ and nvidia/, seven shards each, about 67.6 GB of storage in total, roughly 33.8 GB per hardware tree. The repository was last modified on 24 September. No card rewrite, no benchmark table with numbers in it, no technical report.

The pattern is inverted, and the practical consequence is inverted too. Ming's weights have been runnable since early September by anyone willing to read a config file; its documentation did not exist. InternLumina's documentation existed from day one; its weights did not, and when they arrived they arrived without the comparison tables the card promised.

What each model is actually built to do

Ming-3.0-audio-flash is a native audio-language mixture of experts: 126 billion parameters total, 6.5 billion active per inference step, an audio encoder and projector feeding a 42-layer backbone with 512 experts and 8 active per token, a 157,184-token vocabulary, and a Whisper-style audio tower at 32 layers and width 1,280 with 128 mel bins. Input is audio plus text instructions. Output is text, or text plus synthesised speech through an optional Talker module with built-in Chinese voices and two reference clips in the repository. The declared context is 128K, and the card's serving profile is eight 80 GB H800s in tensor parallel.

InternLumina-U2 is a 16-billion-parameter sparse mixture of experts with 1 billion active, built as a diffusion language model rather than an autoregressive one — it refines a whole sequence in parallel instead of emitting tokens left to right. Its visual representation is fully discrete: eight codebooks of visual tokens from Apple's AToken tokenizer, concatenated into one backbone token per position on the input side and predicted codebook by codebook on the output side. The point of that design is unification: text question answering, image understanding, image editing, text-to-image generation, video understanding and 3D understanding all run through the same vocabulary and the same weights, on a LLaDA-2.0 diffusion backbone.

Read the two descriptions together and the comparison resolves into a single sentence. Ming-3.0-audio-flash is a specialist that is very large and does one modality deeply. InternLumina-U2 is a generalist that is small by total parameters and tries to do five things adequately. Neither design is a compromise of the other; they are answers to different questions.

Turning each one into tokens, which is the real difference

If you only read the parameter counts, you would conclude that Ming is eight times the model. That reading is wrong in both directions, and the reason is the interface each one exposes.

• Input — Ming-3.0-audio-flash: audio and text. InternLumina-U2: text, images, video frames and 3D assets.

• Output — Ming-3.0-audio-flash: text, optionally speech. InternLumina-U2: text, images, edited images; no audio, no video, no 3D generation.

• Decoding — Ming-3.0-audio-flash: autoregressive over a 157,184-token vocabulary. InternLumina-U2: block diffusion over a discrete multi-codebook visual vocabulary.

• Total versus active parameters — Ming-3.0-audio-flash: 126B total, 6.5B active. InternLumina-U2: 16B total, 1B active.

• Weights on disk — Ming-3.0-audio-flash: roughly 259 GB across 25 shards plus a Talker and VAE. InternLumina-U2: 67.6 GB in the repository, split into two hardware trees of seven shards and about 33.8 GB each.

• Licence — both MIT-adjacent in spirit, different in fact: Ming-3.0-audio-flash is MIT per its card, InternLumina-U2 is Apache-2.0 per its repository and card.

• Independent evaluation — none for either.

Hardware, which is where the two diverge hardest

InternLumina-U2's most interesting engineering claim is not the architecture; it is the backend story. The repository ships both an Ascend tree and an NVIDIA tree, with the lab stating that training, evaluation and inference were adapted to both Huawei NPUs and NVIDIA GPUs with operator-level numerical alignment between them. For anyone whose procurement runs through non-NVIDIA accelerators, that is a real differentiator and it is checkable in the file tree: two complete checkpoint directories, each with its own config, tokenizer and parallel plan.

Ming-3.0-audio-flash has no such story. Its reference profile is eight 80 GB H800s; its validated alternative is a four-GPU SGLang profile that returns text only. There is no NPU path in the repository, and the tiny sibling at 9.3B total and 2.7B active is the answer for anyone who cannot commit a node.

Both models are, in the contemporary sense, deployment projects rather than pip installs. InternLumina's own materials are explicit that the generation path is not standard: the block-diffusion generate API is not the Transformers generate interface, the AToken tokenizer wants FlashAttention, and the 3D pipeline expects Blender 4.5 to render GLB assets into views before tokenisation. Ming's path is more conventional — AutoProcessor plus AutoModelForCausalLM with trust_remote_code=True — but the memory floor is eight times higher.

What the vendor numbers say, and what they cannot say

Neither table can be compared to the other, because the two models do not share a single benchmark. That is not a dodge; it is the honest shape of the evidence.

On inclusionAI's evaluation images, Ming-3.0-audio-flash reports a Chinese ASR average of 3.45 error rate, an English average of 3.69, a dialect average of 9.40, MMAR 74.40 and BigBenchAudio 99.40. The same images show where it loses: AMI diarization error 25.29 against 11.16 for MiMo-V2.6 Pro, and a long-audio Chain score of 12.31 against 34.54 for Qwen3.8-Omni Flash. Its text benchmarks are the weakest section of the card — GPQA Diamond 82.45, MMLU-Pro 80.00, IFBench 59.00 — which is what a 6.5B-active audio specialist looks like when it is asked to reason in prose.

InternLumina-U2's card reports ChartQA 86.52, CharXiv-DQ 83.65, MathVision 33.22, HallusionBench 62.15, VideoMME 51.26, MVBench 59.74, 3D MM-Vet 41.8, DPG-Bench 87.10 and GenEval 0.81 — and then undercuts itself, because the project page's own table shows LLaDA2.0-Uni at GenEval 0.89, better than the model that claims to surpass it. The card's headline caveat, "Preliminary, partial results. Full comparison tables will appear in the upcoming technical report," is still true today. There is no technical report.

Read together, the two sets of numbers describe complementary strengths and one shared weakness: neither has been reproduced by anyone outside the lab that trained it.

Which one, for which reader

If your product involves voice — meeting transcription with dialect coverage, call-centre audio, spoken answers, dubbing — Ming-3.0-audio-flash is the only one of the two that applies, and honestly the only open-weight model in its class with published dialect numbers this specific. Plan on a node, or start with the tiny sibling and accept its weaker long-audio scores.

If your product involves images — document reading, chart QA, editing, 3D asset description, or a single model that both reads and draws — InternLumina-U2 is the only one that applies, and the Ascend tree is a genuine reason to look if you are not buying NVIDIA. Plan on fighting the tokenizer.

If you need both, you are running two models, and the interesting question becomes operational: two checkpoints, two toolchains, two sets of licence terms, one multimodal product. That is the normal state of open-weight work in late 2026, and it is why routing exists.

The routing layer, stated narrowly

OrcaRouter serves neither checkpoint. Ming-3.0-audio-flash and InternLumina-U2 are both open-weight models you run yourself; neither is in our catalogue and neither will appear in a call to our API. What we do provide is the surrounding text work: the prompt that frames an image-edit instruction, the summariser that turns a Ming transcript into structured notes, the classifier that decides which of the two models an incoming asset should go to.

That last piece is the one worth naming. A pipeline that handles both voice and vision ends up with a routing decision — audio goes to the Ming node, images to the InternLumina node — and that decision is a model call like any other. Putting it behind one API for more than 200 models at 0% markup, where the provider list price passes straight through, keeps the dispatcher on the same key as everything else. Automatic failover matters here too, not for the self-hosted checkpoints but for the classification step that feeds them: if the router's model is degraded, the fallback should be a different model, not a broken pipeline.

None of that makes a 259 GB checkpoint into a hosted service, and it would be misleading to suggest otherwise.

What to watch next

For Ming-3.0-audio-flash, the missing pieces are a licence file in the repository (MIT currently lives in card metadata) and a serving integration for the audio-language path; vLLM-Omni supports the previous Ming generation and has an open pull request for the 7.9B T​TS model, but nothing merged for the 126B reader. For InternLumina-U2, the missing piece is the technical report and the full tables it promises; until then the GenEval row against LLaDA2.0-Uni is the only cross-model comparison anyone can check, and it does not favour the newer model.

The thing both releases have in common is more useful than either gap. Two capable labs published substantial open-weight multimodal models and told nobody. That is not a sign the models are bad — it is a sign that the release announcement has become optional, and that the repository is now the primary source. Read it.