Shopify's post-trained open model beats a frontier model while cutting serving costs 90%

NVIDIA Blog · rss · 2026-10-22

NVIDIA rounded up European retail deployments of open models and physical AI:

Shopify post-trained Nemotron 3.5 Lightning with Megatron Bridge and NeMo RL as the GraphQL subagent in Sidekick; on its internal benchmark it beats a leading frontier model in accuracy while cutting serving costs 90% and responding 7x faster. vLLM optimization brought 4.5x speedups and up to 12x lower inference cost across five more use cases.
Zooplus rebuilt product discovery with 86% search accuracy; Photoroom standardizes 40,000 listing images a day for Vestiaire Collective, lifting liquidity 5%; Omi's ProductDrop anchors generated images to digital twins (used by L'Oréal, whose ingredient screening shrank from months to hours with NVIDIA ALCHEMI).
Physical AI: Otto Group simulates warehouse ops in digital twins at its largest fulfillment center; ALLSIDES builds sub-millimeter 3D digital twins for adidas, Nike and Zara; CATCHES' virtual try-on lifted conversion 55% and add-to-cart 80%.

More from Companies & People