0%experts made useless during training
Traditional mixture-of-experts training lets 20% to 56% of experts stop receiving tasks and become dead weight — capacity that's paid for and never recovered. ACE measures zero wasted experts, across every scale tested.
Measured results
| Metric | Traditional model | ACE |
|---|---|---|
| Experts that become useless during training | 20%–56% | 0% |
| Memory use for the same capacity | baseline | up to 2× lower |
| Final quality (same training budget) | baseline | equal or better, consistent across every scale tested |
Confirmed across multiple model scales and multiple independent runs — the zero-waste pattern repeats consistently.
ACE is integrated with DAS — growing the model's knowledge with a new domain now costs a fraction of what it used to, keeping the same exact retention guarantee.
See DAS →ACE is one of the pieces behind Mouhn AV — the AI architecture we're building for genuine scientific reasoning.
Meet Mouhn AV →In practice, every expert the model trains actually gets used — so the same training budget buys more real capability, not silent waste.