Physix FrontierPhysix Frontier · News Briefing Card (Leiphone · Nov 2, 2026)
Meituan releases 1.6T-parameter LongCat-2.5 model
KEY FACTS
- Meituan has released LongCat-2.5, a next-generation large model built for long-horizon agentic tasks.
- The model is natively multimodal, with 1.6 trillion total parameters.
- LongCat-2.5 was trained entirely on domestic compute, making it the largest training run ever carried out on domestic compute.
KEY DATA
1.6 trillionTotal parameters
PHYSIX OBSERVATION
By running a 1.6-trillion-parameter model end to end on domestic compute, Meituan, if the claim holds, would demonstrate that domestic chips can support ultra-large-scale training, a strong signal for domestic-chip/domestic-model co-design; combining long-horizon agents with multimodality may also point toward more complex task automation.
Source: Leiphone report