Meituan releases LongCat-2.5 for multimodal agent tasks

Availability and evidence

Meituan’s API changelog marks LongCat-2.5 available on November 2, following the September 25 preview. The company newsroom explicitly gives the same release day and says users can access the model through chat and the API. [1][2]

The release combines visual understanding with agentic coding and extended engineering workflows. Meituan describes 1.6 trillion total parameters and training on domestic computing hardware. TechNode reports that GitHub and Hugging Face releases were still planned, so hosted access does not establish downloadable weights or an open license. [2][3]