Meituan Unveils LongCat-Next AI Model with Unified Vision and Speech Architecture

On April 3, the MiTi team officially launched the native multimodal large model LongCat-Next. This model moves beyond the conventional "language foundation plus plugins" approach by converting images, audio, and text into a unified stream of discrete tokens. This allows the AI to natively "see" and "hear" the physical world, processing these inputs just as it does text.

Technical Core: DiNA Architecture Enables "Modality Internalization"

To eliminate barriers between different data types, MiTi developed the DiNA (Discrete Native Autoregressive) architecture, achieving a deep unification in multimodal modeling:

Complete Modality Unification: The model uses the same parameters, attention mechanisms, and loss functions for text, images, and audio.

Symmetry of Understanding and Generation: Within a single mathematical framework, predicting the next text token constitutes "understanding," while predicting an image token is "generation." Both processes show significant synergistic benefits during training.

Extreme Compression: Utilizing the dNaViT Visual Tokenizer, it handles inputs at any resolution. Through an 8-layer residual vector quantization process, it achieves up to 28x compression in pixel space while preserving critical details for tasks like OCR and financial document analysis.

Empirical Performance: Discrete Modeling Has No Inherent Limits

LongCat-Next delivers performance that surpasses specialized models across multiple benchmarks, effectively challenging the traditional notion that "discretization inevitably causes information loss":

Fine-Grained Perception: On the OmniDocBench for dense text scenarios, it outperforms not only Qwen3-Omni but also the specialized vision model Qwen3-VL.

Visual Reasoning: It scored an impressive 83.1 on MathVista, demonstrating robust, industry-grade logical reasoning.

Cross-Modal Collaboration: While maintaining leading language capabilities (C-Eval 86.80), it supports low-latency parallel generation of text and speech, along with customizable voice cloning.

Industry Insight: A Foundation for Physical-World AI

Large language models have long been centered on text. The breakthrough of LongCat-Next is its proof that physical-world information can be discretized and modeled like language. When an AI possesses a unified "native language," it becomes more intelligent and intuitive when using tools, writing code, or interpreting complex charts.

MiTi has now open-sourced the LongCat-Next model and the dNaViT tokenizer. This efficient, high-potential native discrete architecture provides developers with essential tools for building AI that can perceive and interact with the real world.