Meituan Releases LongCat-2.5: A 1.6-Trillion-Parameter Multimodal Coding Model with Open Weights

Meituan released the open weights of LongCat-2.5, a 1.6-trillion-parameter multimodal coding model, under the permissive MIT license on November 9, 2026, making it available on Hugging Face for unrestricted commercial use.

LongCat-2.5: Capabilities and Architecture

LongCat-2.5 is designed as a mixture-of-experts model, distinguishing itself through its multimodal capabilities. It can process various input types, including screenshots and video, in addition to traditional text. This allows the model to understand visual context alongside code, enabling it to write or revise software through multiple iterative steps. Such functionality positions LongCat-2.5 as a versatile tool for complex coding and agent tasks.

Performance Benchmarks and Comparisons

The new model demonstrates substantial performance improvements over its predecessor, LongCat-2.0, across several key benchmarks:

• Terminal Bench v2.1: LongCat-2.5 scored 88.4, a notable increase from LongCat-2.0's 70.8.
• DeepSWE v1.1: LongCat-2.5 achieved 64.6, significantly outperforming LongCat-2.0's 9.7.
• SWE Atlas: LongCat-2.5 recorded 58.3, compared to LongCat-2.0's 33.1.

In direct comparisons with other leading models, LongCat-2.5 shows competitive results:

• On Terminal Bench v2.1, LongCat-2.5 is within one point of top-tier models like GPT-5.6 Sol and Claude Opus 5.
• However, on the more recent Terminal Bench v4.0, LongCat-2.5 scored 38.5, trailing Claude Opus 5, which achieved 53.9.

These benchmarks highlight LongCat-2.5's strong performance in specific coding and reasoning tasks, while also indicating areas where other models currently hold an advantage.

Open-Source Access and Commercial Implications

The decision by Meituan to release LongCat-2.5 under the MIT License is a critical aspect of this announcement. The MIT License permits broad use, modification, and distribution, including for commercial purposes, without significant restrictions. This open-source approach can foster wider adoption and community-driven development, potentially leading to new applications and integrations.

For those seeking hosted solutions, Meituan also offers discounted API access to LongCat-2.5 through its longcat.ai ai platform. This provides an alternative for users who prefer not to host the model themselves, combining the benefits of open weights with managed service convenience.

Conclusion

Meituan's release of LongCat-2.5 with open weights under the MIT License on November 9, 2026, represents a notable contribution to the AI news landscape. Its multimodal capabilities and strong benchmark performance, particularly in coding and agent tasks, position it as a significant tool for developers and researchers. The open-source nature of the release, coupled with available hosted API access, aims to encourage broad adoption and further innovation in AI-driven software development.

Sources

• meituan-longcat/LongCat-2.5 · Hugging Face
• models.dev/models/meituan/longcat-2.5-preview.toml at dev · anomalyco/models.dev
• Update LongCat-2.5 model card and documentation assets · meituan-longcat/LongCat-2.5 at f3ae0ad
• meituan-longcat/LongCat-Video-Avatar · Hugging Face
• meituan-longcat/LongCat-2.0 · Hugging Face