InternLumina-U2 (NVIDIA checkpoint)
Shanghai AI Laboratory · 2026-09-24
The NVIDIA-trained checkpoint of Shanghai AI Laboratory's unified diffusion model for text, image, video and 3D understanding, image generation and editing. Its weights became available in the official repository on September 24, 2026; it shares the LLaDA2 and eight-codebook AToken architecture with the separately trained Ascend checkpoint. The advertised 16B-A1B backbone count does not establish an exact total for the complete multimodal system.
Benchmark scores
| Benchmark | Score |
|---|
No benchmark scores recorded.