LongCat-Flash-Lite
68.5B-total, ~3B-active MoE where over 30B parameters are N-gram embeddings, offered as a better scaling axis than adding experts.
Meituan reports specific regimes where scaling an embedding table beats adding MoE experts on the Pareto frontier, with lower I/O pressure. 256K context via YaRN; MIT. LongCat-2.0 later inherits the idea (135B N-gram embedding parameters).
- Date
- Wednesday, 28 January 2026
- Lab
- Meituan (LongCat)
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Total / active parameters | 68.5B / ~3B | company |
HF repo created 2026-01-27 UTC; weight files appear in commits from 2026-01-30 and the tech report from 2026-02-06.
Sources
This record was checked and corrected against its sources on 6 October 2026. How we check