Nvidia posted a new multimodal model and told nobody about it.
A repository called PixelUMM appeared on Hugging Face on 1 October with a 15.2-billion-parameter checkpoint. Nvidia published no paper and no announcement, and the files are raw training shards nobody can load. The card says the model reads and generates both images and video, and is built on Alibaba's Qwen3-8B.
Why it mattersNvidia built this on a Chinese open base model, which shows how far those weights travel. Downloads still sit at zero, so treat it as a lab artefact rather than a product.
Read in the app — free, in 9 languages
Related stories
A court threw out a sentence because the victim spoke by AI.2026-10-04You can now tell an image model exactly where to put things.
2026-10-03Cloudflare gave away two models that only choose, never chat.
2026-10-03Germany released a free model built to think in German.
2026-10-03Google gave its strongest new model first to cyber defenders.
2026-10-01