LongCat-2.0 Goes Sparse: 1.6T-Parameter MoE, Only 48B Active per Token

LongCat-2.0 is a Mixture-of-Experts (MoE) language model built at very large scale, carrying 1.6 trillion total parameters but routing only 48 billion of them as “active” parameters on any given forward pass. That ratio is the whole point: the model holds an enormous pool of specialized experts, yet a routing layer engages just a small fraction for each token, so inference cost tracks the active count rather than the full parameter total.

The design reflects where frontier model building has converged. Dense models pay for every parameter on every token, which makes trillion-scale dense networks impractical to serve. Sparse MoE architectures decouple capacity from compute — you can grow total knowledge by adding experts while keeping the per-token compute, and therefore latency and serving cost, roughly fixed. A ~33x gap between total and active parameters puts LongCat-2.0 firmly in the aggressively sparse camp.

For anyone evaluating it, the headline numbers set expectations but leave the operative questions open: how the router balances load across experts, what the real memory footprint is to host all 1.6T parameters, and how it benchmarks against comparable sparse models. Those details — not the raw parameter count — will determine whether the efficiency tradeoff actually pays off in practice.