InternLumina-U2 Preview: One 16B Diffusion Model for Image, Video and 3D — Weights Still Coming

At 04:03 UTC on 1 September 2026, Shanghai AI Laboratory created the GitHub repository InternLumina-U2. Thirteen minutes later the same organisation registered an identically named repository on Hugging Face. There was no launch post, no model card and no announcement of any kind — just the inference code for InternLumina-U2, a 16B-parameter mixture-of-experts model (1B active) that the lab bills as one model covering image understanding, text-to-image generation, image editing, video understanding and 3D understanding through a single discrete-token interface. The code is real and public; the model itself is not — yet. The weights are marked "coming soon", the Hugging Face repository is an empty placeholder, and the promised technical report has not appeared. What went out quietly on September 1 is nevertheless the most complete public picture of this model that exists anywhere.

If this is the first you are hearing of InternLumina-U2, that is the point. Nothing about the release was announced, and no independent coverage appears to exist yet — early-signal write-ups are exactly where a model like this surfaces first, before the launch post and sometimes before the weights. Everything below is drawn from the GitHub repository, the project page and the Hugging Face stub the lab itself published on September 1, and anything beyond those sources is explicitly labeled as inference.

What actually shipped on September 1

The GitHub repository InternLM/InternLumina-U2 was created on 1 September 2026 and saw three commits that day — the init, a pass marking the model link "coming soon" and the benchmark results "preliminary", and a navigation fix. It is licensed Apache-2.0 and carries a README titled "Intern Lumina U2: A Multi-Codebook Diffusion Large Language Model for Omni-Visual Understanding, Image Generation and Editing", an English and a Chinese version, a full inference harness, and a roadmap. One quirk worth noting: the repository's own news entry is stamped "[2026-08]", yet the init commit and both repository timestamps are dated 1 September 2026 — treat early September as the actual release date, not August. The repository below is the entire public release: every file visible in the tree is part of what shipped, and nothing more exists anywhere yet.

• GitHub — InternLM/InternLumina-U2, created 2026-09-01 under Apache-2.0, with inference code for text, text-to-image, image understanding, image editing, video understanding and 3D understanding.

• Hugging Face — internlm/InternLumina-U2, created 2026-09-01, currently holds only a .gitattributes file and a 28-byte README whose entire content is the Apache-2.0 license header. No weights, no config, no tokenizer files.

• Project page — internlm.github.io/InternLumina-U2, with architecture figures, a preliminary benchmark table and placeholder demos; the technical report is marked "coming soon".

The inference code is organised as one driver script with a section per task: plain text generation, text-to-image at up to 1024×1024 with configurable CFG and timesteps, image understanding over natural images and dense charts, instruction-based image editing with an optional dual-CFG mode, video understanding over 64 sampled frames, and 3D understanding over GLB/GLTF assets that are rendered into 64 color-and-depth views through a bundled Blender pipeline before tokenization. That breadth of tasks is the whole design thesis of the model, and it is worth pausing on before diving into the architecture.

One model, six jobs, one token vocabulary

InternLumina-U2 belongs to a fast-moving line of research betting that language and vision should share a single discrete-token interface rather than live in separate understanding and generation stacks. The lab's own earlier work in this direction — Lumina-DiMOO, the unified discrete-diffusion model shown at WAIC 2025 — is cited alongside LLaDA2.0-Uni, Show-o2 and MMaDA as the pioneering systems InternLumina-U2 builds on and claims to surpass. The specific bet InternLumina-U2 makes is that the bottleneck for these unified models is not the architecture but the visual vocabulary: a single codebook caps how much information one visual token can carry, while discrete–continuous hybrids that split understanding and generation across different representations force the model to maintain two stacks anyway.

InternLumina-U2's answer is fully-discrete, multi-codebook modeling. The visual side uses Apple's AToken tokenizer, which describes each visual position with eight complementary codebooks — an attempt to preserve local texture, embedded text, geometry and high-frequency detail without inflating sequence length. On the input side, each of the eight codebooks keeps its own embedding, and the eight per-position embeddings are concatenated and projected into a single backbone token. On the output side, prediction is spatially parallel but codebook-depth autoregressive: the diffusion backbone denoises many masked spatial positions in parallel, and a multi-codebook autoregressive head then predicts the eight codes of each position in order.

• Scale — 16B total parameters, 1B active (16B-A1B), a sparse MoE.

• Language backbone — LLaDA-2.0 MoE diffusion language model, whose block-diffusion objective is extended to visual tasks through joint multi-task training (vendor description).

• Visual representation — 8-codebook fully-discrete tokens from Apple's AToken tokenizer.

• Hardware — adapted to both NVIDIA GPUs and Huawei Ascend NPUs across training, evaluation and inference, with operator-level numerical alignment between backends.

What that buys in practice, if the claims hold, is the oldest dream of the multimodal field: a single set of weights that can read an image, edit it, draw a new one from text, answer questions about a video, and describe a 3D asset — with understanding and generation reinforcing each other through the same interface rather than being stitched together from specialist models. That is also the claim most worth a skeptical eye, because it is the hardest one to make true.

The numbers, such as they are

Every benchmark InternLumina-U2 has is vendor-reported, preliminary and partial. The README and project page say so themselves: "Preliminary, partial results. Full comparison tables will appear in the upcoming technical report." No independent evaluation exists, because the weights are not public. Treat the figures below as the lab's own claims, not verified scores.

• Chart / document understanding — ChartQA 86.52, CharXiv-DQ 83.65 (vendor-reported).

• Visual math reasoning — MathVision 33.22, DynaMath 56.37; the lab says this beats the understanding-only InternVL3-8B on both.

• Hallucination robustness — HallusionBench 62.15.

• Video understanding — VideoMME 51.26, MVBench 59.74, MLVU 57.03 (project page).

• 3D understanding — 3D MM-Vet 41.8.

• Text-to-image — DPG-Bench 87.10, TIIF-Bench Short/Long 84.60/85.95, GenEval 0.81 (project page).

• Image editing — ImgEdit 3.83.

The comparison story is where an honest reader should slow down. The README claims InternLumina-U2 surpasses unified models such as InternVL-U, LLaDA2.0-Uni, Show-o2 and Lumina-DiMOO on fine-grained understanding and generation benchmarks — but the project page's own table shows InternLumina-U2 at GenEval 0.81 against LLaDA2.0-Uni's 0.89, so the model trails at least one competitor on at least one headline generation metric. Picking a few favorable numbers from a partial table is exactly what the "full comparison tables in the tech report" caveat is designed to soften. The numbers are a directional signal from the lab, no more.

What is real versus what is promised

The release scope is unusually explicit, and it matters for anyone deciding whether they can try this today. Shipped: the inference code. Not shipped: the weights, the training code (promised in a separate repository), the Ascend training-and-inference stack, and the technical report. The Hugging Face repository that will eventually hold the model is a stub — the screenshot below is the entire current contents of the page a reader reaches when they click the "Model (coming soon)" link in the README.

Even the released code cannot produce output yet. The inference driver expects a split Hugging Face checkpoint directory and the AToken tokenizer weights at a specific path — neither of which is in the repository — so what is downloadable today is a specification of how the model will run, not a running model. And when the weights do land, self-hosting will not be a routine pip-install job. The model uses a block-diffusion generate API rather than the standard Transformers generate interface, the AToken tokenizer requires FlashAttention, the code needs a visible GPU at import time, the root driver is single-process, and the 3D pipeline expects Blender 4.5 for asset encoding. That is a real deployment project, not a weekend experiment — which is exactly the kind of friction a routing layer exists to absorb for most teams.

When the weights land, treat it like the unproven model it is

No one can run InternLumina-U2 today, and no provider can serve it, because the weights do not exist in public. That will change at some point — Apache-2.0 licensing and the lab's 2026 pattern of aggressive open-sourcing (the ArchSpace "open everything" push, InternVL3.5, the Intern-S2 science models) make a weight drop likely rather than hypothetical. When it happens, the rational first step is not to bet a production pipeline on a day-one diffusion model with unusual serving requirements and zero independent evaluations. It is to run it side-by-side with the model you already trust, on a test path, with automatic failover to the proven model the moment the new one misbehaves. That is the use case a routing layer is built for: one API across 200+ models, provider list prices passed through at 0% markup, and failover that turns "try the new thing" into a routing rule rather than a migration. OrcaRouter is set up that way — and to be clear, we do not route InternLumina-U2 and cannot, because the weights are not released. This section is about the workflow for when they are, not a claim that the model is available anywhere today.

What to watch next

Four events would move InternLumina-U2 from preview to reality, in rough order of likelihood. First, the Hugging Face repository populating: safetensors files, a config, and the AToken tokenizer weights appearing in that stub is the moment the model actually exists. Second, the technical report, which is supposed to carry the full comparison tables and would turn the partial vendor numbers above into something checkable. Third, the training-code repository and the Ascend stack, which matter for anyone who wants to fine-tune or run this on non-NVIDIA hardware. Fourth, a first independent evaluation — an arena Elo, a reproduced ChartQA or GenEval run — that tests the "one model, six jobs" promise without the lab's own selection bias. The code being public on September 1 means the architecture is already knowable; the weights being absent means everything about actual quality is still a claim. Watch the model repository rather than the announcement feed — the interesting parts are already public, and the model itself is probably not far behind.