One-sentence summary
Key specs from Meituan's Nov 2 LongCat-2.5 launch: 1.6T total params, ~48B active per token, native multimodal, 1M-token context, trained entirely on domestic compute.
About the Data
This table compiles model specs Meituan disclosed at the November 2, 2026 LongCat-2.5 launch, based on official release and mainstream tech coverage.
Core Data
| model | Metric | Value |
|---|---|---|
| LongCat-2.5 | Total parameters | 1.6T |
| LongCat-2.5 | Active params per token | ~48B |
| LongCat-2.5 | Context length | 1M tokens |
| LongCat-2.5 | Modality | Native multimodal (incl. image understanding) |
| LongCat-2.5 | Training compute | Fully domestic compute |
| LongCat-2.5 | First release | LongCat-2.5-Preview |
Field Notes
- Total parameters: full model size.
- Active params per token: parameters activated per token in the MoE design.
- Context length: max tokens handled in one pass.
- Training compute: where training compute came from.
Sources
- Meituan: Meituan launches LongCat-2.5
- Huxiu: Meituan launches 1.6T LongCat-2.5