LongCat‑2.5 combines native visual tokens with long-horizon agent tasks

What is changing?

Meituan has released LongCat‑2.5, a natively multimodal model designed for long-horizon agentic tasks. According to the official technical post, its mixture-of-experts architecture has 1.6 trillion total parameters, with about 48 billion activated per token. The model combines visual understanding and agentic coding in one backbone.

LongCat‑2.5 converts images into multi-level discrete tokens at native resolution. A fixed visual tokenizer allows these sequences to be precomputed and cached, while the backbone processes text and visual inputs in a shared token format. Training uses long task trajectories, verifiable examples derived from agent failures, and large-scale long-horizon reinforcement learning.

The LongCat API changelog says the model became available on November 2, 2026, targeting visual coding, creative design, 3D tasks, and complex project development. Compatibility with agent harnesses including Claude Code, Codex, and Hermes is also listed. Provider benchmark claims should not be treated as independent evaluation.