❯ Meituan releases LongCat-2.5 under MIT for visual and multistep coding tasks

Downloadable weightsMeituan published the LongCat-2.5 weights on November 9 after introducing the model on November 2. The MIT license permits commercial use, modification and redistribution subject to conditions such as retaining the license notice. The model accepts text and visual inputs and is designed for jobs that take multiple steps.

From images to actionThe model can read screenshots, charts and video. A developer could give it a screenshot of a broken page, have it locate the relevant code and then inspect the rendered result after a change. In Meituan’s tests, LongCat-2.5 scored 88.4 on Terminal Bench v2.1, up from 70.8 for its predecessor, and 64.6 on DeepSWE v1.1, up from 9.7.

Hosted access or self-deploymentLongCat-2.5 has 1.6 trillion total parameters and activates about 48 billion per token on average. It uses only part of the model for a given input, though a self-hosted deployment still needs hardware to hold the full weights. Meituan’s hosted API lets teams trial the model by usage before committing hardware to a private deployment.

▮ SIGNALOpen weights give teams handling large volumes of screenshots and internal code a self-hosting option, with the 1.6-trillion-parameter footprint setting the hardware cost of that control.