Meituan announced today the release of LongCat-2.5 weights under the MIT license on Hugging Face. The model, launched on November 2, 2026, has 1.6 trillion total parameters in an MoE architecture, about 48 billion active per token, native multimodal support (text, image, and video), and a 1 million token context window.
What happened?
The announcement came from the official @Meituan_LongCat account on X. The weights are available on Hugging Face, allowing developers to download and run the model locally. Unlike version 2.0, which was already open-weight, 2.5 was initially offered only via API, and the weights were released one week after general availability.
Why does this matter?
LongCat-2.5 is optimized for long-horizon agent tasks such as complex code development, visual coding, design, and 3D tasks. It integrates native vision into the backbone via multi-level discrete tokens. Internal benchmarks show strong performance on Terminal Bench and BrowseComp, competing with frontier models. The open-source release under MIT lowers barriers for research and use in AI and game projects.
What changes in practice?
Developers can now run the model without relying solely on Meituan's API (pricing around $0.30/$1.20 per million tokens). This benefits open-source AI communities, especially on local hardware or Chinese clusters. Technical details include support for OpenAI and Anthropic compatible protocols, and long-range RL training.
Sources: official Meituan LongCat announcement and technical reports.
By Grok
