The Meituan LongCat team released LongCat-2.5 on November 1, 2026. It has about 1.6 trillion total parameters with roughly 48 billion activated per token. Building on LongCat-2.0, it adds stronger native visual understanding and agentic coding, aimed at long, multi-step tasks such as full project builds.
According to the team, LongCat-2.5 represents images as multi-level discrete tokens so the model handles visual and text input in one unified format. Pairing that with coding skills gives it strong visual coding, meaning it can turn what it sees into working web pages, and it also supports creative design and 3D tasks. Showcased demos include a real-time fluid simulation page, 3D digital exhibitions and city dioramas.
In the official evaluation table, LongCat-2.5 scores 78.6 on Vision2Web and 88.4 on Terminal Bench v2.1, with several coding benchmarks run using Claude Code as the harness. The team notes that some scores for other models come from public reports while others were self-run and are indicative only.
For training, the team scaled long-horizon reinforcement learning for end-to-end engineering and converted real agent failure trajectories into verifiable training tasks. On the inference side, it built 4-bit quantized serving on domestic AI accelerators, reporting roughly 43 percent higher prefill and decode throughput than its 8-bit baseline.