LongCat-2.5
Meituan · 2026-11-02
LongCat-2.5 is Meituan's natively multimodal mixture-of-experts model, marketed as 1.6 trillion total parameters with about 48 billion activated per token, which takes images and video as multi-level discrete visual tokens. Meituan says it was trained entirely on domestic AI accelerators and serves it with W4A4 inference through the LongCat API, with a 1M-token context and a thinking switch; weights are announced as coming soon. It is distinct from the API-only LongCat-2.5-Preview of September 25, 2026.