Observed arrival · 2026-11-05
expXa: a V100-focused MoE inference engine
A research-led inference engine tuned for quantized mixture-of-experts models running on NVIDIA V100 GPUs.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- Developers running quantized MoE models on V100s
- Worth noticing
- The 35B comparison uses three runs per engine; 30B CPU-layer results come from separate sessions and should not be compared across rows.
Field notes
The implementation focuses much of its work on hybrid execution: it ports PXA 4’s CPU expert queue and quantized tiles, adds parallel activation preparation, and can switch tile modes between requests. The GPU work includes a Volta tensor-core attention schedule and a two-token Q6_K expert kernel; the page also links to individual studies and a data section.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
◎NicheUnusually specific use
One card from the complete issue