Meituan Releases LongCat-2.5 Multimodal MoE Model

Performance & Benchmarks

LongCat-2.5 delivers a massive jump in software engineering capabilities. On the DeepSWE v1.1 benchmark, the model scored 64.6, a staggering improvement over LongCat-2.0's score of 9.7. This performance places it in the same elite tier as GPT-5.5 (67.0) and Claude Opus 5 (68.8), effectively closing the gap between open-weights models and top-tier proprietary systems.

In codebase navigation and question-answering, the model achieved a score of 58.3 on the SWE Atlas Codebase QnA benchmark. This matches the performance of GLM-5.3 and outperforms both kimi-k3 (54.3) and DeepSeek V4.1 Flash (55.2). These scores demonstrate the model's ability to maintain coherence and accuracy when interacting with massive, multi-file repositories.

The evaluation methodology focused on 'agentic' success rates—measuring not just whether the model could write a snippet of code, but whether it could successfully complete a multi-step task, such as debugging a broken UI component by analyzing a screenshot and then modifying the underlying CSS and React components.

• DeepSWE v1.1: 64.6 (vs GPT-5.5's 67.0)
• SWE Atlas Codebase QnA: 58.3 (vs DeepSeek V4.1 Flash's 55.2)
• Primary Benchmark Focus: Agentic coding and visual-to-code reasoning

API Pricing

Meituan has released the weights under a permissive MIT license, making it highly accessible for local deployment. For those preferring managed services, pricing is available through various domestic and international aggregators.

Currently, for the hosted versions of the LongCat family, users can expect competitive pricing similar to other high-parameter MoE models. However, because the weights are non-gated and free on ModelScope, the most cost-effective path for developers is self-hosting the W4A4 quantized variant.

• Self-hosting: Free (MIT License)
• Quantized Variant: W4A4 available for low-VRAM environments
• Managed API: N/A (Check ModelScope/Hugging Face for provider-specific endpoints)

Use Cases

The most compelling use case for LongCat-2.5 is 'Visual-to-Code' agentic workflows. Imagine a developer providing a screenshot of a broken checkout flow; LongCat-2.5 can identify the visual discrepancy, locate the relevant component in a massive codebase, and propose a fix. This makes it ideal for autonomous UI/UX testing and front-end development agents.

Beyond coding, the model's ability to handle 3D tasks and creative design through visual-semantic alignment opens doors for automated CAD assistance and digital asset management. Its strength in long-horizon tasks also makes it a prime candidate for complex RAG (Retrieval-Augmented Generation) pipelines where visual documents (like blueprints or complex charts) must be parsed alongside massive text corpora.

• Autonomous Coding Agents (Full-project builds)
• Visual Debugging & UI/UX Automation
• 3D Task Reasoning & Creative Design
• Complex Multimodal RAG

Getting Started

Developers can access LongCat-2.5 immediately via Hugging Face and ModelScope. For production environments with limited hardware, Meituan provides a W4A4 quantized version specifically optimized for faster inference and lower memory footprints.

To get started, we recommend cloning the repository from ModelScope to take advantage of the optimized loading scripts designed for domestic AI accelerators. For those integrating into existing agentic frameworks (like LangChain or AutoGPT), the model's discrete visual tokenization makes it highly compatible with standard multimodal prompt structures.

• Weights: Available on Hugging Face & ModelScope
• Optimized Version: W4A4 Quantized variant
• Deployment: Optimized for domestic AI accelerators

Sources