LongCat-2.5 Packs 1.6 Trillion Parameters to Turn Visuals into Code
LongCat-2.5: Native Multimodality for Agentic Tasks
LongCat-2.5 is a native multimodal model built for long-horizon agentic tasks, with 1.6 trillion total parameters and roughly 48 billion activated per token. It encodes images as multi-level discrete tokens so one backbone handles vision and text in a unified format, and large-scale long-horizon reinforcement learning sharpens end-to-end coding across complex codebases and full-project builds. The team also reports W4A4 inference and long-context training on domestic AI accelerators.
We also convert real-world agent failure trajectories into reproducible, verifiable training tasks, continually adding targeted training data to address model capability gaps.
- slopnt
The barplots are being miscolored, so only their own model is visible. Its hard to compare them. It highly suggests that the blog post was never looked at by a human.
- BugsJustFindMe
> "the charts below compare it with GPT-5.5, GPT-5.6-Sol, Claude-Opus-4.8, and Claude-Opus-5"
There should be a rule against releasing seemingly-obsolete comparisons without explaining why you're doing it. If you don't give a cost reason, it makes you look like you're trying to hide something.
I'm not saying that the given comparison is inherently bad. I'm saying that they failed to include any useful-to-potential-users reason for why those specific non-current model versions are meaningful.
- colingauvin
For API, this seems competitive based on the released benchmarks. For running at home, this is pointless when DS4.1 is more or less equivalent but half the total parameters and less than half the active parameters.