InternLumina-U2 vs North Micro Vision Instruct: A Doc Reader You Can Run Today vs a 16B Model That Isn't There Yet

If you need a model that reads documents, screenshots and charts this week, this comparison has a short practical answer: North Micro Vision Instruct is a 2.4-billion-parameter open-weight model from Cohere, released under Apache-2.0, downloadable since 10 August 2026, and good enough at document tasks to be worth pointing at your own files today. InternLumina-U2 is Shanghai AI Laboratory's 16-billion-parameter diffusion model — a far more ambitious design that also claims chart and document understanding, among a half-dozen other jobs — but its weights are not public, its Hugging Face repository is an empty stub, and nobody anywhere can run it yet. The longer answer is about what happens when two Apache-2.0 models make overlapping claims about reading, but only one of them is a thing you can hold in your hand.

The 2.4B that actually exists

North Micro Vision Instruct is the vision specialist in Cohere's North family: roughly 2.4B total parameters, a compact model built to read at native resolution. The design point is that most vision models downscale images to a fixed grid and lose the fine detail that documents live on — so North Micro Vision Instruct accepts images at their native resolution up to about an A4 page at 200 dpi, preserving aspect ratio and small text. Cohere's reported numbers are strong for the size class: DocVQA at 0.921, ChartQA at 0.808, OCRBench at 0.792, all Cohere-reported and unreproduced, with a validated multimodal context around 8K tokens and a documented weak spot at MMMU (0.329) — a reminder that it is a reading specialist, not a reasoning generalist. It has no hosted API of its own and no standard hosted route; the practical story is self-hosting a 2.4B model, which is small enough to run on a single modern workstation GPU. Community adoption has been steady — the Hugging Face card was showing tens of thousands of downloads a month by late August.

The 16B that is a promise

InternLumina-U2 is a different kind of artifact. What Shanghai AI Laboratory published on 1 September 2026 is inference code and an architecture, not a model: a sparse mixture-of-experts diffusion LLM, 16B total parameters with 1B active, built around a fully-discrete eight-codebook visual representation. Among the six task families the repo's driver script covers — text, image understanding, text-to-image, image editing, video understanding, 3D understanding — image understanding explicitly includes dense charts and documents. The project page's preliminary table lists chart and document figures the lab reports in the same range the specialist claims: ChartQA 86.52, CharXiv-DQ 83.65, alongside a HallusionBench 62.15 and a MathVision 33.22. The page calls the results "preliminary, partial," the technical report that would carry the full tables is marked coming soon, and — the decisive fact — there are no weights with which anyone could check.

The scoreboard

Every figure below is vendor-reported. North Micro Vision Instruct's are Cohere's own evaluation; InternLumina-U2's are the lab's preliminary partial table.

• Size and shape — North Micro Vision Instruct: ~2.4B dense, native-resolution reader. InternLumina-U2: 16B total / 1B active sparse MoE, discrete multi-codebook diffusion model.

• What it does — North Micro Vision Instruct: read images, documents, charts and UI screens; answer in text, with grounding. InternLumina-U2: claims reading plus generation — understand images/video/3D, generate and edit images.

• Weights — North Micro Vision Instruct: public since 10 August 2026 (CohereLabs/North-Micro-Vision-Instruct, Apache-2.0). InternLumina-U2: not public; Hugging Face stub empty, weights marked "coming soon."

• Headline reading scores — North Micro Vision Instruct: DocVQA 0.921, ChartQA 0.808, OCRBench 0.792 (Cohere-reported). InternLumina-U2: ChartQA 86.52, CharXiv-DQ 83.65, HallusionBench 62.15 (lab-reported, preliminary).

• Document context — North Micro Vision Instruct: native resolution up to ~A4 at 200 dpi, ~8K validated multimodal context. InternLumina-U2: dense charts and natural images handled via the AToken multi-codebook encoder; context not yet specified.

• Known weaknesses — North Micro Vision Instruct: MMMU 0.329, no tool calling — a reader, not a reasoner or an agent. InternLumina-U2: trails at least one named rival on GenEval (0.81 vs 0.89) in its own partial table; everything unverified.

• How you run it — North Micro Vision Instruct: self-host a 2.4B model; vLLM support was still landing when this was written. InternLumina-U2: no one can run it — no weights, and the released driver needs a checkpoint directory and tokenizer files that are not in the repo.

The same benchmark, two very different epistemics

ChartQA is the cleanest way to see the difference between the two claims. Both models report a ChartQA number; North Micro Vision Instruct reports 0.808 as an accuracy fraction, and InternLumina-U2 reports 86.52 as a percentage — the same benchmark, essentially the same result in the high-80s-to-mid-80s band on different scales, give or take the different evaluation splits and prompt setups that make cross-paper ChartQA comparisons treacherous even when both models exist. The epistemic difference is what matters: North Micro Vision Instruct's 0.808 is attached to a downloadable artifact, so any team with a GPU and a set of charts can reproduce or refute it this week. InternLumina-U2's 86.52 is attached to a roadmap. A number you cannot check is a number you should not build on — which is exactly the position the lab itself acknowledges by labeling its table preliminary.

The CohereLabs/North-Micro-Vision-Instruct Hugging Face card, captured 27 August 2026 — the 2.4B native-resolution vision-language description, the Apache-2.0 license, and Cohere's reported document-reading benchmarks.

Reading, versus reading and drawing

The architectural philosophy gap is worth more than the benchmark gap. North Micro Vision Instruct is a narrow specialist: it reads, and it does that one job affordably enough that you can run it on every page, every screenshot, every invoice in a pipeline without watching your GPU bill. That cheapness is the feature — document intelligence at scale is a cost problem as much as an accuracy problem. InternLumina-U2 is the opposite bet: a huge sparse model trying to fold reading together with image generation, image editing, video understanding and 3D understanding into one shared discrete-token interface, on the theory that understanding and generation reinforce each other. If it works, it is a different class of tool — but "if it works" is doing a lot of work, and a 16B-A1B diffusion MoE with a custom tokenizer and Blender-rendered 3D preprocessing will never be the cheap always-on reader that a 2.4B specialist is. These are not really substitutes. They are a tool you can deploy today and a research direction you can prepare for.

The InternLM/InternLumina-U2 GitHub repository, captured 2 September 2026 — the repo landing page with the "Omni-Visual Understanding, Image Generation and Editing" description and the per-task inference scripts; the image-understanding path among them is what targets natural images and dense charts.

What this means if you are building a document pipeline

Neither model is hosted on OrcaRouter — North Micro Vision Instruct is a self-host model with no hosted API, and InternLumina-U2 has no weights for anyone to host — so the routing angle here is not "call them both through one key today." It is the pattern you would use the day either of them changes: a document pipeline almost always pairs a cheap reader with a bigger reasoner, and a routing layer's value is expressing that pairing as one call and keeping the 0% markup pass-through honest on the hosted models you do use. When an Apache-2.0 checkpoint like InternLumina-U2 finally lands, the sensible move is not to rip out a working document stack — it is to put the newcomer on a test path behind automatic failover, so it earns production traffic by surviving it, and the moment it fails on a real chart the call falls back to the model that has already proven itself. One API across 200+ models is the mechanism; failover is the safety net; the specialist you can run today is what pays the bills while the 16B promise is still being kept.

The verdict

For document and chart reading in the here and now, North Micro Vision Instruct is the only one of the two that exists in a usable sense — small, Apache-2.0, downloadable, with strong vendor-reported scores you can actually go check. InternLumina-U2 is the more interesting long-term artifact, because it is trying to make reading one skill among six in a single unified model, and if that works it changes what "vision model" means. Watch its empty Hugging Face repo the way you would watch a launch countdown: the day safetensors files appear, the promise becomes a model, and the comparison becomes a real one. Until then, do not let a vendor-reported 86.52 on a chart benchmark outrank a downloadable 0.808 in your decision-making.