You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
SPA Mode. Frameworks that support a full SPA mode not just a route
caching repeated navigations is also a big performance thing, it might be nice for the benchmark to include a repeated navigation somewhere to demonstrate and measure the reuse of cached data (from TanStack Team)
Add a navigation-after-hover benchmark: hover for 100 ms, click, and measure time from click to destination content painted across all frameworks. (from TanStack Team)
It would be interesting to test hydrated components with server rendered markup in addition to the current tests for client rendering. Client rendering in React for example should be largely equivalent across different frameworks that support React, so the tests are likely mostly measuring React rather than the framework. The bigger differences may be where the frameworks provide unique takes on the server–client boundary to optimize performance (islands vs React Server Components etc.)
a cpu throttled variant, everything runs unthrottled desktop right now which hides js cost. something like 4x would show where hydration actually hurts
hmr latency next to dev server startup, the time from a file edit until the browser shows it. startup is measured but the edit loop is what devs feel all day
My biggest request for that list would be this: measure an actual navigation. I think this is probably one of the most important metrics for a "meta-framework benchmark". Basically: the time from a click on a link until the destination of that link is rendered.
Currently the only measure close to this is the INP (called "Interaction Breakdown" in https://frameworks.e18e.dev/all-frameworks) but this stops before the actual navigation, it's basically just the click handling itself, and since all frameworks included in this benchmark have async operations between the click and the render of the next page, it doesn't actually measure the navigation.
I just realized that this benchmark does not measure the amount of bytes loaded by the client. There is "body size" which is the HTML document in bytes, but another good one would be "total bytes on initial load" (i.e. HTML + JS bundles since i don't think these apps ship any CSS)
Sometimes a package will include dependencies as part of their distribution, without including it in package.json#dependencies. For example, such packages may not appear via npm ls, and using npm audit won't reveal vulnerabilities of vendored packages such as this vercel/next.js#83459.
It would be great to have a deterministic way to reveal the number of vendored dependencies each framework has as part of the benchmark, and to represent them in some way as part of the total dependency count
One DevTime test idea: record build memory high-water and constrained-memory pass/fail alongside build time. On Linux, a dedicated cgroup's memory.peak and OOM events would be more useful than a single process's max RSS when a build creates multiple workers. Fix CPU/RAM and cache conditions across frameworks, distinguish an empty build directory from a repeat build, and alternate repeat blocks to expose drift.
We ran a small synthetic Next.js 16.4.6 application in one local ARM64 VM, stopped/resized in A/B/A order (2 vCPU/4 GiB, 4 vCPU/8 GiB, 2 vCPU/4 GiB). Three clean-directory runs per block gave medians 10.422 → 3.223 → 3.928 seconds; repeat builds with .next retained gave 4.050 → 1.313 → 1.412 seconds. All 18 builds completed with no guest OOM kills. Our memory observation was only sampled guest-wide MemAvailable, not process-tree or cgroup peak memory. We could not isolate CPU vs RAM, cache or sequential host effects, and the second A block substantially changed the interpretation. This is one fixture, not a framework ranking. A standardized memory metric would make the result more useful than build time alone.
Disclosure: MOTSELL Engineering (VPS provider) ran this local test; AI assisted with automation and drafting. No commercial benchmark or sales link.
Activity
delucis commentedon Jul 14, 2026
It would be interesting to test hydrated components with server rendered markup in addition to the current tests for client rendering. Client rendering in React for example should be largely equivalent across different frameworks that support React, so the tests are likely mostly measuring React rather than the framework. The bigger differences may be where the frameworks provide unique takes on the server–client boundary to optimize performance (islands vs React Server Components etc.)
dreyfus92 commentedon Sep 28, 2026
nice list, two things i would add are:
Sheraff commentedon Oct 5, 2026
My biggest request for that list would be this: measure an actual navigation. I think this is probably one of the most important metrics for a "meta-framework benchmark". Basically: the time from a click on a link until the destination of that link is rendered.
Currently the only measure close to this is the INP (called "Interaction Breakdown" in https://frameworks.e18e.dev/all-frameworks) but this stops before the actual navigation, it's basically just the click handling itself, and since all frameworks included in this benchmark have async operations between the click and the render of the next page, it doesn't actually measure the navigation.
delucis commentedon Oct 5, 2026
This might be relevant here as presumably @Sheraff means measuring SPA navigation: https://developer.chrome.com/docs/web-platform/soft-navigations
Sheraff commentedon Oct 22, 2026
I just realized that this benchmark does not measure the amount of bytes loaded by the client. There is "body size" which is the HTML document in bytes, but another good one would be "total bytes on initial load" (i.e. HTML + JS bundles since i don't think these apps ship any CSS)
AlexanderKaran commentedon Nov 11, 2026
Most of the meta-frameworks have a built in Image component or an official lib for it. Taking a look at these across a few tests would be good
AlexanderKaran commentedon Nov 14, 2026
#847
From @theetrain
motsell commentedon Nov 20, 2026
One DevTime test idea: record build memory high-water and constrained-memory pass/fail alongside build time. On Linux, a dedicated cgroup's
memory.peakand OOM events would be more useful than a single process's max RSS when a build creates multiple workers. Fix CPU/RAM and cache conditions across frameworks, distinguish an empty build directory from a repeat build, and alternate repeat blocks to expose drift.We ran a small synthetic Next.js 16.4.6 application in one local ARM64 VM, stopped/resized in A/B/A order (2 vCPU/4 GiB, 4 vCPU/8 GiB, 2 vCPU/4 GiB). Three clean-directory runs per block gave medians 10.422 → 3.223 → 3.928 seconds; repeat builds with
.nextretained gave 4.050 → 1.313 → 1.412 seconds. All 18 builds completed with no guest OOM kills. Our memory observation was only sampled guest-wideMemAvailable, not process-tree or cgroup peak memory. We could not isolate CPU vs RAM, cache or sequential host effects, and the second A block substantially changed the interpretation. This is one fixture, not a framework ranking. A standardized memory metric would make the result more useful than build time alone.Disclosure: MOTSELL Engineering (VPS provider) ran this local test; AI assisted with automation and drafting. No commercial benchmark or sales link.