InternLumina-U2 vs Microsoft Mage-VL: Two Quiet Labs Betting That Vision Tokens Should Carry More

Microsoft Mage-VL and InternLumina-U2 both shipped the way research labs ship when they are not sure the result is a product yet: silently. Microsoft published the Mage-VL weights and code on Hugging Face on 25 July 2026 with no announcement; Shanghai AI Laboratory's InternLM organization published the InternLumina-U2 inference code on GitHub on 1 September 2026 with no announcement either. Five weeks apart, two of the largest labs in the world dropped models that share a quiet conviction — that the way we turn vision into tokens is the bottleneck — and then left them for the community to notice. One of them you can download and run today, awkwardly. The other you cannot run at all, because its weights do not exist yet. This is the honest map of the two, and of why "both vendor-reported, both unproven" is not the same sentence for both of them.

Five weeks, two bets on representation efficiency

The throughline is real, not rhetorical. Mage-VL is a codec-native video model: instead of decoding a stream into uniformly-sampled frames and pushing a dense grid of patches through a frozen web-pretrained vision transformer, it follows the structure of modern video codecs, separating a stream into anchor frames and the compressed motion data between them, and ingesting that compact representation directly. Microsoft's reported payoff is a large reduction in visual tokens and a meaningfully faster streaming perception path — the company frames it as fixing a kind of Moravec's paradox for video language models, where small real-time perception jobs cost as much compute as hard offline reasoning. Mage-VL is a watcher: it consumes video efficiently and answers questions about what it sees, in near-real time.

InternLumina-U2 is a bet on the same bottleneck from the other direction. Where Mage-VL compresses video by following the codec, InternLumina-U2 compresses every visual input by describing each position with eight complementary discrete codes instead of one, using a tokenizer the lab calls AToken. The wager is that a single codebook caps how much information one visual token can carry, so a multi-codebook vocabulary lets one 16B-parameter diffusion model do understanding and generation through the same interface — read a chart, answer a video question, describe a 3D asset, generate an image from text, edit one on instruction — without a separate generation stack. Mage-VL wants to perceive streams cheaply; InternLumina-U2 wants to perceive and draw with one set of weights.

The scoreboard

Both models' figures are vendor-reported and unreproduced, so the scoreboard labels every row with its provenance.

• Shape — Mage-VL: a compact codec-native vision-language model (roughly 5B parameters in BF16) wrapped around a Qwen text backbone, trained for image and video understanding. InternLumina-U2: a 16B-parameter MoE with 1B active (16B-A1B), a sparse diffusion LLM covering understanding, generation and editing.

• Visual representation — Mage-VL: codec-native tokens following video-codec structure (anchor plus motion); reported >75% visual-token reduction on streaming video. InternLumina-U2: fully-discrete eight-codebook tokens from the AToken tokenizer.

• What it does — Mage-VL: watch image and video input, answer in text; built for streaming perception. InternLumina-U2: claims text, image understanding, text-to-image, image editing, video understanding and 3D understanding.

• Weights — Mage-VL: public on Hugging Face since 25 July 2026 (microsoft/Mage-VL, Apache-2.0). InternLumina-U2: not public — the Hugging Face repo is an empty stub, weights marked "coming soon."

• Headline numbers — Mage-VL: Video-MME 64.0, NExT-QA 83.1, OVO-Bench 64.0, all Microsoft-reported. InternLumina-U2: ChartQA 86.52, MathVision 33.22, VideoMME 51.26, GenEval 0.81, all lab-reported, preliminary and partial.

• Serving reality — Mage-VL: downloadable, but there is no standard vLLM or SGLang path; the code requires trust_remote_code, and no inference provider currently deploys it. InternLumina-U2: not servable by anyone — the weights do not exist in public, and even the released inference driver expects checkpoint files that are not in the repository.

"Available but awkward" is not the same as "not available"

This is the distinction that matters most if you are actually trying to build something. Mage-VL is real in the way that counts first: the weights are on Hugging Face, the card has seen heavy download traffic since late July, and a small ecosystem of community quantizations has sprung up around it. "Real" does not mean "easy." The model card shows a custom-code tag, the visual encoder is not the standard frozen ViT that existing serving stacks assume, and Microsoft's own repo carries the practical consequence — running it means trust_remote_code and no turnkey provider deployment. But a determined team with a GPU can load it, point it at a video stream, and see whether the codec-native thesis holds for their data. That is an enormous difference from InternLumina-U2, where the same sentence ends differently: a determined team can read the inference code, and then it stops, because there are no weights to load.

The microsoft/Mage-VL Hugging Face card, captured 27 August 2026 — the codec-native streaming description, the Apache-2.0 license, the arxiv tag, and an inference-provider section showing no provider currently deploys the model.

The benchmark asymmetry follows the same line. Both models' numbers are vendor claims with no independent reruns yet. But Mage-VL's claims at least sit on top of a downloadable artifact, which means the gap between claim and verification is a matter of someone running it. InternLumina-U2's claims sit on top of a roadmap item — "full comparison tables will appear in the upcoming technical report" — and no artifact exists that could verify them. The lab's own partial table even shows the model trailing at least one rival it says it surpasses (GenEval 0.81 against LLaDA2.0-Uni's 0.89), which is a useful reminder to read its "preliminary, partial" label literally.

Where they could compose instead of compete

The most useful way to read these two is as adjacent rungs on a ladder, not as rivals for the same job. A codec-native watcher like Mage-VL is interesting precisely because it is cheap enough to run continuously — the thing that watches every frame of a stream and only escalates when something worth a bigger model's attention appears. The bigger model it escalates to could, in principle, be a unified model of the kind InternLumina-U2 claims to be: the always-on gate that decides what to look at, and the deep model that actually reads the chart, follows the 3D scene, or produces the edited image. Both are unproven — Mage-VL unproven-but-runnable, InternLumina-U2 unproven-and-unreleased — so the honest framing is a composition you could build once each side has been independently validated, not a head-to-head you can run today.

The InternLM/InternLumina-U2 GitHub repository, captured 2 September 2026 — the repo landing page showing the multi-codebook diffusion description, the Apache-2.0 license badge, and the inference entry-point scripts that make up the public release while weights stay out of the tree.

What a routing layer does with unproven checkpoints

Neither of these models is hosted on OrcaRouter, and we are not going to imply otherwise — Mage-VL has no standard serving path anywhere, and InternLumina-U2 has no weights. What the routing DSL is genuinely useful for in this situation is expressing the composition above as one call instead of bespoke glue: a cheap always-on perception model feeding a deeper model, chained so that the expensive model only runs when the cheap one says something changed. And for the unproven-checkpoint problem — which is the entire risk profile of both of these — automatic failover is the mechanism that lets a team try a model like Mage-VL on a real path without betting the path on it: the first time the new checkpoint stalls, returns garbage, or throws, the call falls through to a proven model, and no engineer has to be woken up to flip a switch. One API across 200+ models, provider list price passed through at 0% markup, and the new thing auditioned behind a safety net rather than in front of production.

The bottom line

If you want to test the codec-native-video thesis today, only Mage-VL exists — and "exists" comes with caveats about custom code and DIY serving. If you want to test the multi-codebook unified-model thesis, you cannot, because InternLumina-U2 is still a specification waiting on its weights; what you can do is read its architecture now so you are ready the moment the stub on Hugging Face fills in. Treat the Microsoft model as an awkward-but-real artifact and the Shanghai AI Lab model as a well-documented promise, and remember that in this matchup "vendor-reported and unreproduced" applies to both — it just means something different when there is a file you can download.