SSyncDev

InternLumina-U2

SyncDev model directory · Profile updated Nov 19, 2026

Compare with another model

InternLumina-U2 is Shanghai AI Lab's masked-diffusion MoE that understands images, video and 3D and generates and edits images with 8-codebook visual tokens.

Key specifications

Specification score
19.6
License
Apache-2.0

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: Open weights under Apache 2.0; no hosted price.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input
  • Image input
  • Video input

Produces

  • Text output
  • Image output

Capabilities

VisionOpen weights

What is InternLumina-U2?

Unified models usually split the work: discrete tokens for text, continuous latents for images, and two training objectives that pull against each other. InternLumina-U2, Shanghai AI Laboratory's "Intern Lumina U2", keeps everything discrete. Each visual position is described by codes from eight complementary codebooks, and a masked diffusion language model handles text and images in one space.

The payoff is breadth. One backbone does text QA, text-to-image, image understanding, instruction-based editing, video understanding and 3D understanding.

Size: two numbers in circulation

Shanghai AI Lab rounds it to a 16B-A1B model in one place and gives 16.9B total with about 1.4B active in another, alongside a full architecture breakdown. Use the more precise figures for capacity planning.

Component Detail
Repository internlm/InternLumina-U2 (nvidia/ and ascend/ subfolders)
Backbone LLaDA2.0-mini masked diffusion MoE
Layers 20 (1 dense, 19 MoE)
Experts 256 routed + 1 shared, 8 active
Visual tokens AToken, 8 codebooks per position
Inputs text, image, video, 3D
Outputs text, image
Licence Apache 2.0

How generation works

At the input, the eight per-codebook embeddings for a position are concatenated and projected into one backbone token, so sequence length does not grow eightfold. At the output, the diffusion backbone denoises many masked positions in parallel while a small autoregressive head predicts each position's eight codes in order. An optional, training-free sampler called PHV speeds image generation by 4.45x, per the authors.

Checkpoints exist for both Huawei Ascend and NVIDIA hardware. The GitHub repository ships the NVIDIA path, plus SFT training code.

Publisher-reported scores

  • GenEval 89, DPG-Bench 87.79, TIIF-Bench (long) 87.1
  • ChartQA 86.2, CharXiv descriptive 85.9, HallusionBench 64.6
  • MathVista 68.0, MMMU 52.2
  • Video-MME 52.6, MVBench 61.8
  • ImgEdit 3.82, GEdit-Bench EN 6.06

Generation and chart reading are the strong suits. General knowledge (MMMU) is modest, which fits a backbone with roughly a billion and a half active parameters.

Timeline

Inference code arrived in August and the repository on 1 September 2026, weights followed in October, and the full report is dated 14 October. Check which revision you pull.

Should you try it?

It is worth a look if you want one small-active-parameter model for both chart understanding and image generation. Generate fifty images from your own prompt set and run fifty chart questions from your documents through it; compare the images with a dedicated generator such as Seedream 5.0 Flash and the charts with SenseNova-U1.5-8B-MoT. Similar models are listed under any-to-any.

How to evaluate InternLumina-U2

Compare the model on the work you need to complete.

Tasks and inputs

InternLumina-U2 is listed for Any-to-any, Image-text-to-text, Text-to-image. It accepts text, image, video and produces text, image. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Choose the output size and quality your design needs. Include drafts and revisions when budgeting an accepted image.

Choosing a deployment

The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

Shanghai AI Lab gives 16.9B total with about 1.4B active, also rounded to 16B-A1B.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

Liquid AI

LFM3-Omni-2B

LFM3-Omni-2B is Liquid AI's open any-to-any model: image, video, audio and text in; text and speech out, with multimodal tool calls on-device.

— context

NVIDIA

Nemotron-Labs-Diffusion-Uni-8B

NLD-Uni (Nemotron-Labs-Diffusion-Uni-8B) is NVIDIA's open diffusion-language model that both understands images and generates them, for research.

— context

NVIDIA

PixelUMM

PixelUMM is NVIDIA's encoder-free 15B unified model that understands and generates text, images and video directly in pixel space. Non-commercial.

— context

sensetime

SenseNova-U1.5-8B-MoT

SenseNova-U1.5-8B-MoT is SenseTime's open unified model for native 4K image generation and instruction-based image editing, Apache-2.0 licensed.

— context

Tencent

Hy-Embodied-RxBrain-1.0

Hy-Embodied-RxBrain-1.0 is Tencent's 6.2B embodied model that reasons in text and imagines goal images, for planning, world-state prediction and video QA.

— context

google

google/gemma-4-E2B-it

google/gemma-4-E2B-it is a source-linked Hugging Face repository assigned to Any-to-any. Its recorded task interface accepts text, image, audio, video and produces text, image, audio, video.

131.07K context