Meituan Releases LongCat Image Open Weights Image Generation Model

Introduction

On December 1, 2025, Meituan officially announced its entry into the generative media space with the release of LongCat Image. Following the success of the LongCat-2.5 multimodal LLM series, this new model marks a strategic pivot for the tech giant, moving from pure language and vision understanding into high-fidelity image synthesis.

This release is particularly significant for the developer community because Meituan has opted to release the model weights openly on Hugging Face. By providing open access, Meituan is positioning the LongCat family as a versatile ecosystem capable of handling both complex reasoning and high-quality visual creation.

• Release Date: December 1, 2025
• Provider: Meituan
• Availability: Open weights on Hugging Face
• Category: Image Generation Model

Key Features & Architecture

LongCat Image is built upon the architectural foundations established by the LongCat-2.5 multimodal lineage. While the LLM variants focus on massive parameter counts for reasoning, LongCat Image is optimized for diffusion-based synthesis and prompt adherence. It leverages advanced latent diffusion techniques to ensure that complex, multi-subject prompts are rendered with high spatial accuracy.

The model is designed to work seamlessly within the broader LongCat ecosystem. This means developers building agentic workflows can transition from a LongCat-2.5 vision-language model (which analyzes a scene) to LongCat Image (which generates a new one) without significant architectural friction. The integration of semantic understanding from the LLM family allows for superior prompt following compared to many standalone diffusion models.

• Architecture: Optimized Latent Diffusion
• Ecosystem: Fully integrated with LongCat-2.5 multimodal models
• Deployment: Open weights for local fine-tuning and research
• Optimization: High prompt adherence for complex semantic instructions

Performance & Benchmarks

In terms of raw aesthetic and semantic quality, LongCat Image has demonstrated impressive results in early testing. Most notably, the model achieved an Artificial Analysis Elo score of 860, placing it in a highly competitive tier alongside industry leaders in the image generation space. This score reflects its ability to satisfy human preference across various styles, from photorealism to digital art.

The evaluation methodology utilized by Artificial Analysis involves a blind side-by-side comparison where human raters judge the alignment of generated images with complex text prompts. This Elo-based approach provides a more holistic view of model 'intelligence' and 'creativity' than traditional pixel-based metrics, ensuring that the model excels in the nuances of composition and lighting.