A controlled load test published on June 9 puts hard numbers behind a claim premium merchants have heard for years. In a Hyvä vs Luma comparison run on identical AWS infrastructure, the Hyvä storefront served median requests 5.2 times faster than the same Magento install on Luma. The headline is real. The more useful story sits in the numbers nobody screenshots.
What the benchmark measured
The Mage-OS benchmark, run by David Lambauer, is refreshingly specific about its setup, which is what makes it worth reading. Both stacks ran Magento 2.4.7 on four m6i.2xlarge nodes, eight vCPU and 32 GB each, on dedicated EKS node groups. Same catalogue: 5,000 products across 100 categories, generated from Magento's own fixture tool so the dataset was identical on both sides. The test drove 250 virtual users through a realistic traffic mix, roughly 30 percent browse, a quarter search, a quarter product detail, a fifth add-to-cart. Luma ran its usual RequireJS, KnockoutJS and jQuery payload. Hyvä ran Alpine.js, Tailwind and Typesense.
The median result is the line that travels: 201 ms on Hyvä against 1,040 ms on Luma. Throughput moved from 36 to 51 requests per second, a 42 percent lift, 40,021 total requests against 28,147 over the run. Hyvä also cut database work to 43 SQL queries per request against Luma's 56. On a warm, well provisioned cluster, the new frontend stack does measurably less work for the same page.
The number the headline hides
Read past the median and the picture gets honest. At the 95th percentile, Luma sat around 8,000 ms and Hyvä around 3,680 ms. Hyvä is still more than twice as fast, but 3.7 seconds at p95 is not a fast storefront. It is a storefront under stress. The reason is buried in the setup: the team deliberately disabled Full Page Cache and Varnish. They wanted to measure the application and the frontend, not the cache layer that every serious production Magento runs in front of it.
That choice is the right call for a benchmark and the wrong number to quote to a client. In production, Full Page Cache and Varnish absorb most catalogue and category traffic before PHP ever wakes up. What the uncached test actually exposes is the backend ceiling both stacks share. Peak CPU never climbed past 25 percent on either side. The bottleneck was PHP-FPM worker concurrency, not raw compute. Hyvä does not fix that. It is the same Magento application underneath. What Hyvä changes is everything in front of it.
Why the frontend payload still wins
For a premium studio shipping commerce in Southeast Asia, the median number is the one that maps to real buyer experience, because real buyers hit cached pages on mid-range Android phones over uneven mobile networks. That is where Luma's RequireJS and jQuery payload costs you. Every kilobyte of blocking JavaScript is parsed and executed on a device with a fraction of an m6i.2xlarge's headroom. The 5.2x median gap on warm infrastructure understates the gap on a budget handset in Jakarta or Manila.
This is the case we make on nearly every Hyvä build. The platform debate, which we lay out in our Shopify Plus vs Magento Hyvä breakdown, is rarely about raw backend throughput. It is about what reaches the browser. Hyvä's smaller footprint, Tailwind's atomic CSS and a search layer like Typesense that answers in tens of milliseconds compound into a storefront that feels instant where it counts, on the device the customer actually holds.
What we would actually change this quarter
Three moves, none of them a re-platform.
First, stop quoting uncached benchmarks to stakeholders. Run your own test with Full Page Cache and Varnish on, the way the site actually serves, then run it again with cache off to find your true backend ceiling. The gap between those two runs is your real cache dependency, and it is worth knowing before a flash sale, not during one.
Second, if you are on Luma and the p95 on cached category pages still drags on mobile, the frontend is your cheapest win. A Hyvä migration is a frontend project, not a backend rebuild, and it leaves your integrations and data model intact. That is why it ships in weeks, not quarters, and why the risk profile is nothing like a platform change.
Third, watch PHP-FPM worker counts and database query volume per request, because that is the ceiling no theme can lift. The benchmark's 56 versus 43 query gap is a reminder that frontend choices ripple into backend load. Fewer round trips per page buys you headroom under the concurrency a regional launch in Southeast Asia will throw at you on day one.