NVIDIA's Gated DeltaNet-2 splits erase and write gates in linear attention, aiming for cleaner memory updates and stronger benchmarks without attention costs.
Some results have been hidden because they may be inaccessible to you
Show inaccessible results