LongCat released LongCat-2.5, a 1.6 trillion parameter mixture-of-experts model. It adds native visual understanding to its agentic coding. The model uses multi-level discrete tokens to unify visual and text inputs. It scales long-horizon reinforcement learning for complex code and full-project builds. Infrastructure improvements target long-context training and inference on domestic AI accelerators. Numerical consistency between training and inference is preserved. Long-context rollout and W4A4 inference are optimized. The model is available via its chat interface and API. GitHub and Hugging Face releases are coming soon.
What changed
LongCat-2.0 lacked native visual understanding and long-horizon agentic reinforcement learning at this scale.
What it unlocks
Running long-horizon agentic tasks with native visual and coding capabilities in one model.
- 1.6 trillion total parameters
- 48 billion activated per token
- 86% higher per-device throughput in long-sequence SFT
What you need to act on it
- access to LongCat-2.5 via API or platform
- domestic AI accelerator hardware for optimal performance
Sources