build1 publisher
Top-k gating decouples DeepSeek-V3's 671B parameter count from its 37B of per-token compute
A top-k selection inside the gate, described in a 2017 Google Brain paper, is what skips the other experts, and the memory floor still tracks all 671 billion parameters because each one stays resident.
Publishers:dev.to
Reality
- Evidence60
- Adoption38
- Hype gap0
- Incentives20
- Confidence55