LongCat-Next: Meituan Open-Sources Native Multimodal Model — 291 HuggingFace Upvotes, Sees/Paints/Talks in One Framework

arXiv / HuggingFace Daily Papers·high signal

Meituan's LongCat-Next introduces DiNA (Discrete Native Autoregressive), a unified framework that tokenizes text, vision, and audio into a shared discrete space under a single autoregressive objective. The key innovation is dNaViT, a visual transformer doing tokenization at arbitrary resolutions with hierarchical discrete tokens. The team open-sourced the model and tokenizers. 291 HuggingFace upvotes makes it the most-upvoted daily paper this cycle.