Meituan Open-Sources LongCat-Next: 3B Parameter Multi-Modal Model Unifies Text, Visual, and Voice Capabilities

Meituan's LongCat team has open-sourced LongCat-Next, a native multi-modal model based on MoE architecture with 3B activated parameters. It unifies five capabilities—text, visual understanding, image generation, voice understanding, and synthesis—within a single autoregressive framework. The model and its supporting tokenizer are open-sourced under MIT license, with weights available on HuggingFace.

Its core design uses the DiNA (Discrete Native Autoregressive) paradigm: paired tokenizers and decoders for each modality convert visual and audio signals into discrete tokens, sharing the same embedding space with text, and completing all tasks via unified next-token prediction. The visual component dNaViT extracts 'visual words' with dynamic tokenization and decoding, maintaining image generation quality at a 28x compression ratio, especially in text rendering.

Among models with similar activated parameters (A3B), LongCat-Next performs well in benchmarks like visual understanding (MathVision:64.7), image generation (LongText-EN:93.15), programming (SWE-Bench:43.0), and Agent tool calls (Tau2-Retail:73.68). In unified model comparisons, it leads with an MMMU score of 70.6, without sacrificing pure text or Agent capabilities.