LongCat-Image
A 6-billion-parameter bilingual image generation and editing model covering all 8,105 standard Chinese characters, with open-source SOTA on image editing benchmarks.
Independent third-party resource. Not affiliated with or endorsed by LongCat or Meituan.
Overview
LongCat-Image is a 6-billion-parameter bilingual (Chinese-English) foundation model for image generation and editing, released by Meituan's LongCat team in December 2025. Unlike many competing image generation models that use Mixture-of-Experts (MoE) architectures with nearly 20 billion parameters or more, LongCat-Image achieves competitive performance with a compact 6B-parameter design — significantly reducing VRAM requirements and deployment cost. The model is particularly notable for its Chinese text rendering capability: it covers all 8,105 standard Chinese characters as defined by the General Standard Chinese Characters Table, exceeding both major open-source and commercial solutions in character coverage according to the publisher. LongCat-Image sets open-source state-of-the-art results on image editing benchmarks (ImgEdit-Bench 4.50, GEdit-Bench 7.60/7.64) and achieves strong text-to-image generation scores (GenEval 0.87, DPG-Bench 86.8). It is available on LongCat Web and the LongCat App with 24 built-in editing templates.
Architecture
LongCat-Image uses a hybrid MM-DiT (Multi-Modal Diffusion Transformer) and Single-DiT backbone, as described in the technical report (arXiv:2512.07584). The MM-DiT component handles the cross-modal interaction between text prompts and image latents, while the Single-DiT component performs the denoising steps in pixel space. This dual-backbone design allows the model to share a compact parameter budget between understanding (text-to-image alignment) and generation (image quality) without the overhead of MoE routing. The publisher reports that the model was trained with a carefully curated data pipeline combining synthetically captioned images, human-annotated aesthetic data, and bilingual (Chinese-English) text-image pairs. A reward model fine-tuning stage applies RL-based optimization for aesthetic quality, photorealism, and Chinese text rendering accuracy — the last being the most technically challenging requirement given the complexity of Chinese character shapes and the need to render them correctly in generated images.
Training
The model was trained by the Meituan LongCat team using a multi-stage pipeline. Initial training used large-scale bilingual text-image pairs with synthetic captions. A subsequent reward model fine-tuning stage optimized for three objectives: aesthetic quality (human preference scores), photorealism, and Chinese text rendering accuracy. The reward model was trained on human-annotated comparisons of generated images. The full training pipeline, including mid-training and post-training checkpoints, was released open-source — making LongCat-Image one of the most transparently documented image generation models available.
Key Benchmarks
Publisher-reported scores. Not independently verified by this site. Descriptions summarize what each benchmark measures.
| Benchmark | Score | What It Measures |
|---|---|---|
| ChineseWord | 90.7 | Chinese text rendering in generated images — measures accuracy across all 8,105 standard characters defined by the PRC General Standard Chinese Characters Table |
| ImgEdit-Bench | 4.50 | Image editing quality benchmark — evaluates the model's ability to modify specific regions of an image while preserving the rest |
| GEdit-Bench | 7.60 / 7.64 | General image editing benchmark — two sub-scores measuring global editing quality across diverse editing instructions |
| GenEval | 0.87 | Text-to-image generation benchmark — measures compositional generation ability including object counting, attribute binding, and spatial relationships |
| DPG-Bench | 86.8 | Dense Prompt Generation benchmark — evaluates how well the model handles prompts with many detailed constraints simultaneously |
Deployment
LongCat-Image is accessible through the LongCat Web interface (longcat.ai) and the LongCat App (iOS and Android), where it powers the image generation, image editing, and text-to-image features with 24 built-in editing templates. Model weights and training artifacts are available on HuggingFace under the meituan-longcat organization. The compact 6B-parameter design means the model can run on consumer-grade GPUs, unlike the ~20B MoE alternatives that typically require server-grade hardware. The publisher has released mid-training checkpoints, post-training checkpoints, and the full training toolchain, allowing the community to fine-tune or adapt the model for specific image generation tasks.
Capabilities
Described in the cited primary documentation. Not an independent performance evaluation.
- Bilingual (Chinese-English) text-to-image generation.
- Text-driven image editing with open-source SOTA (ImgEdit-Bench 4.50, GEdit-Bench 7.60/7.64).
- Chinese text rendering covering all 8,105 standard characters (ChineseWord 90.7).
- Photorealistic image generation with enhanced aesthetic quality through curated reward models.
- Compact 6B parameter design — significantly smaller than ~20B MoE alternatives — for minimal VRAM usage.
- Text-to-image generation quality: GenEval 0.87, DPG-Bench 86.8.
- Available on LongCat Web and LongCat App with 24 editing templates.
Access & License
Open source. Review the repository license file before commercial use. Multiple model versions released including mid-training and post-training checkpoints.
Sources
- LongCat-Image Technical Report (arXiv:2512.07584)
Primary sourcePublished 2025-12-08Accessed 2026-07-26
Technical report describing the MM-DiT + Single-DiT hybrid architecture, data curation pipeline, RL fine-tuning with reward models, and benchmark results including ChineseWord 90.7 and GenEval 0.87.
- LongCat-Image on HuggingFace
Primary sourceAccessed 2026-07-26
Model weights available under the meituan-longcat organization. Includes mid-training and post-training checkpoints, training toolchain, and multiple model versions for text-to-image and image editing.
- LongCat Image on LongCat Official Website
Publisher documentationAccessed 2026-07-26
LongCat-Image is accessible through the LongCat Web interface and LongCat App, with 24 image editing templates and image-to-image capabilities.
Related comparisons
- LongCat-Image and Seedream 5.0: Open-Weight vs Closed Image Generation Compared
A source-based comparison of LongCat-Image (6B bilingual open-weight model, Chinese text rendering, self-hostable) and Seedream 5.0 Pro (ByteDance's closed flagship, 14-language rendering, layer separation, API-only). Covers architecture, text rendering, editing, pricing, licensing, and deployment.
- LongCat-Image and Qwen-Image 2.0: Two Open-Weight Chinese Image Models Compared
A source-based comparison of LongCat-Image (6B bilingual open-weight model, 8,105-Chinese-character rendering, self-hostable) and Qwen-Image 2.0 (Alibaba's 7B open-weight model, native 2K, unified generation-editing). Covers architecture, text rendering, benchmarks, licensing, and deployment — an open-weight vs open-weight comparison.
- LongCat-Image and Seedream 4.5: Open-Weight vs Closed Image Generation Compared
A source-based comparison of LongCat-Image (6B bilingual open-weight model, 8,105-Chinese-character rendering, self-hostable) and Seedream 4.5 (ByteDance's closed image model, 4K output, 10 reference images, multi-image fusion). Covers text rendering, resolution, editing, pricing, and licensing — the predecessor of Seedream 5.0.
- LongCat-Image and FLUX.2: Open-Weight Image Models Compared
A source-based comparison of LongCat-Image (6B bilingual open-weight model, 8,105-Chinese-character rendering, self-hostable) and FLUX.2 (Black Forest Labs' 32B open-weight image family, 4MP editing, 10 reference images). Covers architecture, text rendering, licensing nuances, pricing, and deployment.
FAQ
- Where can the primary source for LongCat-Image be found?
- The official GitHub repository (meituan-longcat) and HuggingFace page contain the primary documentation, model weights, and license for LongCat-Image.
- Is LongCat-Image available through the LongCat API?
- LongCat-Image 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 2026-07-26. 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: 2026-07-26