One-sentence summary
Using Meituan's LongCat-2.5, tell apart MoE total parameters from per-token active params: 1.6T total does not mean 1.6T compute; the active count is what matters.
Why two numbers matter
A MoE (mixture-of-experts) model splits parameters into expert sub-networks and activates only some per token, so total parameters and actual compute are not equal.
Steps
- Read total parameters first: LongCat-2.5 is 1.6T, showing large capacity headroom.
- Then read active params per token: about 48B, the amount actually used in each forward pass.
- Estimate inference cost from active params, not total.
- When comparing models, compare active params and context length first, then total.
Common Mistake
Treating 1.6T total as the per-token compute overstates inference cost; an MoE's real cost tracks its active parameter scale.
Sources
- Meituan: Meituan launches LongCat-2.5