On November 2, Meituan released LongCat-2.5, a native multimodal model built for long-horizon agent tasks that need sustained execution. Meituan says the model has 1.6 trillion total parameters, with about 48 billion activated per token, and was trained entirely on domestic compute. The substantive change is not the parameter count: the full release adds benchmark scores and inference optimisation for domestic chips, whereas the version that went live on September 25 was only an API preview.
What changed between the preview and the full release
The September preview kept LongCat-2.0's parameter scale, added image input, and published no official benchmarks. The full release keeps the same scale and adds large-scale long-horizon reinforcement learning: training took in more long trajectory data and turned failed runs by real agents into reproducible training tasks. Visual signals are represented as multi-level discrete tokens and processed by the same backbone as text, supporting visual coding, creative design and 3D tasks. Meituan says mid-training used sequences of up to 1 million tokens, with cumulative training running to several trillion tokens, and that W4A4 inference was optimised for domestic chips. In a brand operations analysis case shared by Meituan, the model spotted a growth-quality problem — payment GMV up 23.2% while channel profit fell 1.7% — and proposed resource-allocation advice.
How to read the benchmark scores
Meituan's benchmark table puts LongCat-2.5 at 88.4 on Terminal Bench v2.1, above LongCat-2.0's 70.8, and at 91.9 on BrowseComp. Meituan says several scores fall between Claude Opus 4.8 and Claude Opus 5, with some above the latter. All of these figures come from Meituan and have not been independently verified, and on the same table the model scores only 25.5 on Program Bench, below the comparison set's top score of 37. Reading only the headline scores would overstate it; the full table suggests strength in terminal and browsing tasks rather than in every coding task.
Who can use it now, and where the limits are
LongCat-2.5 is available through Meituan's online service and API platform, works with agent tools such as Claude Code, Codex and Hermes, and has GitHub and Hugging Face releases still planned, with no date given. In other words, developers can run long tasks with it today but cannot download weights to self-host; teams that need self-hosting or offline deployment gain nothing new from this release. Teams that need long-horizon coding and business analysis can first test its stability on real projects through the API, then decide whether to adopt it long term. LongCat-Next's native multimodal approach when it was open-sourced
