LongCat-Next
A discrete native multimodal model unifying text, vision, and audio under a single autoregressive objective via the DiNA framework.
Independent third-party resource. Not affiliated with or endorsed by LongCat or Meituan.
Overview
LongCat-Next is a discrete native multimodal model released by Meituan's LongCat team in March 2026, representing a departure from conventional 'language model + modality adapters' architectures. Instead of routing images and audio through separate vision or speech encoders that feed into a text-centric language model, LongCat-Next maps all modalities — text, vision, and audio — into a single shared discrete token space and trains a unified autoregressive model under a single Next Token Prediction objective. This approach, called DiNA (Discrete Native Autoregressive), treats image-to-text and text-to-image as the same conditional token prediction problem, eliminating the structural fragmentation between understanding and generation pipelines. Built on the LongCat-Flash-Lite MoE backbone (68.5B total, ~3B activated), LongCat-Next achieves competitive performance across visual understanding, image generation, coding, agent tool use, and audio tasks. It achieves 28x visual compression through 8-layer RVQ (Residual Vector Quantization) while maintaining strong generative quality.
Architecture
The DiNA framework consists of two key components described in the technical report (arXiv:2603.27538). First, dNaViT (Discrete Native Any-resolution Visual Transformer) performs tokenization and de-tokenization at arbitrary resolutions without forced resize, crop, or pad — critical for OCR-heavy and chart-heavy workloads where fine visual details matter. dNaViT uses 8-layer RVQ to progressively encode residual information, achieving 28x compression while preserving both semantic concepts and fine-grained visual details. Second, a shared autoregressive backbone (the Flash-Lite MoE) processes text tokens, discrete visual tokens, and discrete audio tokens under a single objective function. The publisher reports that this unified setup reduces the historical conflict between understanding and generation objectives: understanding loss rises only marginally versus pure-understanding training, while generation loss improves versus pure-generation training. MoE routing gradually develops modality specialization while maintaining a shared representation space.
Key Benchmarks
Publisher-reported scores. Not independently verified by this site. Descriptions summarize what each benchmark measures.
| Benchmark | Score | What It Measures |
|---|---|---|
| MMLU-Pro | 77.02 | Challenging variant of MMLU — measures multimodal understanding while maintaining strong text knowledge across 57 subjects |
| C-Eval | 86.80 | Chinese comprehensive evaluation benchmark — 52 subjects across humanities, science, and engineering |
| SWE-Bench | 43.0 | Software engineering benchmark — generates patches for real GitHub issues while maintaining multimodal capabilities |
| τ²-Bench (Retail) | 73.68 | Agentic tool-use benchmark in retail domain — measures real-world tool calling combined with multimodal understanding |
Deployment
LongCat-Next model weights, tokenizers (dNaViT, RVQ, SAE), and inference code are open-sourced on HuggingFace and GitHub (meituan-longcat/LongCat-Next). A public demo is available at longcat.chat/longcat-next. The model benefits from the same deployment advantages as Flash-Lite (~3B activated parameters), making it feasible for self-hosted deployment on standard GPU instances. The release includes tokenizers for community development, allowing researchers to experiment with the DiNA framework for their own multimodal applications.
Capabilities
Described in the cited primary documentation. Not an independent performance evaluation.
- Unified text, vision, and audio under a single autoregressive objective.
- DiNA (Discrete Native Autoregressive) shared discrete space.
- dNaViT for tokenization at arbitrary resolutions.
- 28x visual compression while maintaining generative quality.
- Built on LongCat-Flash-Lite MoE (A3B) backbone.
- Open-source tokenizers for community development.
Access & License
Open source. Model, tokenizers, and code publicly released.
Sources
- LongCat-Next (arXiv:2603.27538)
Primary sourcePublished 2026-03-29Accessed 2026-07-27
DiNA framework + dNaViT. LongCat-Flash-Lite A3B backbone. 28x compression ratio. RVQ + SAE for discrete visual representations.
- LongCat-Next on HuggingFace
Primary sourceAccessed 2026-07-27
Model weights and tokenizers.
- LongCat-Next GitHub Repository
Primary sourceAccessed 2026-07-27
Model implementation, tokenization pipeline, inference code.
FAQ
- Where can the primary source for LongCat-Next be found?
- The official GitHub repository (meituan-longcat) and HuggingFace page contain the primary documentation, model weights, and license for LongCat-Next.
- Is LongCat-Next available through the LongCat API?
- LongCat-Next is documented in the meituan-longcat GitHub organization. Refer to the publisher API documentation at longcat.chat for current API availability.
- Is this page an official LongCat or Meituan page?
- No. This page is published by an independent third-party site. It is not affiliated with, endorsed by, or sponsored by LongCat or Meituan.
- How often is this page updated?
- This page was last verified on 2026-07-27. Content is reviewed when new publisher documentation or model releases become available.
Independent third-party disclosure
This page is published by an independent third-party site. It is not affiliated with, endorsed by, sponsored by, or operated by Meituan, LongCat, or any of their affiliates. The content summarizes publicly-available primary documentation and does not represent the views of any referenced organization.
Last reviewed: 2026-07-27