PagedWeight quantizes MoE expert weights per page

PagedWeight quantizes each expert page independently instead of picking one bit-width for the whole MoE model, per arXiv 2607.16184.

MoE models are usually the efficiency pick over dense ones. That flips once long-context serving pushes them memory-bound, since expert weights and cache compete for the same HBM. For anyone self-hosting Qwen or DeepSeek on a fixed GPU budget, that's the difference between one card and two, a straight 2x cost swing.

The paper's fix is quality-aware, dynamic quantization applied per page rather than a flat bit-width across every expert. That's a narrower move than shrinking the whole model to save memory. It's aimed at freeing room for the KV cache without giving up precision everywhere.

Whether that holds outside the paper's own tests is the open question. If per-page quantization keeps its accuracy edge in other serving stacks, the one-GPU-or-two call for self-hosted Qwen and DeepSeek changes. It moves from a hardware budget line to a config flag. That's the kind of change that shows up in an infra bill before it shows up in a benchmark chart.