@Meituan_LongCat trained it on long-horizon agent trajectories at

> Agentic coding model focused on long-horizon task retention

LongCat-2.5 looks much more focused on the part of agentic coding that actually gets difficult: staying useful across a long project.

@Meituan_LongCat trained it on long-horizon agent trajectories at 256K and 1M sequence lengths, plus large-scale RL for multi-step engineering work.

It’s also natively multimodal, representing visual signals as discrete tokens alongside text, which makes things like visual coding, design, and 3D workflows much more interesting.

I’d be especially curious to try it on an unfamiliar repo where it has to understand the codebase, work from a visual reference, impl