Bun HTTP Framework Benchmark

Compare throughput benchmarks from various JavaScript HTTP framework

Prerequistes

• bombardier
• Nodejs
• Deno
• Bun

Run Test

bun benchmark

To run only specific targets, pass their runtime-qualified names:

bun benchmark bun/elysia node/effect

To select frameworks and optionally cap requests per second:

bun benchmark --interactive

To build and verify every supported framework without running a load test:

bun run verify

The verifier starts frameworks one at a time and checks the three core routes, the streamed video, and all background routes. Targets in the benchmark blacklist are skipped.

Select frameworks with the arrow keys and Space, then press Enter and enter an RPS limit. Use 0 or leave it blank for unlimited throughput. Results include the emitted minified bundle size, startup time, and server RSS memory as before / after MB in one column. Deno reports n/a for bundle size because it runs directly.

Dump result will be available at results/[benchmark-name].txt

Each run updates results/results.md and generates README.md from README.template.md with the results and uname -a output. Edit the template to change the README. A selected-target run replaces the table with those targets. If no target produces results, the run preserves README.md.

All Bun and Node targets are minified with Bun.build before their server startup timer begins.

Effect HTTP v4 beta is included for both Bun and Node using one shared router and the official adapter for each runtime.

Elysia AOT variants are included for both Bun and Node.

Deno targets run directly because Bun.build has no Deno target.

Test method: Average throughput

Every implementation registers the same deterministic background routes before the measured routes. These routes are not requested during the load test; they make route lookup and startup closer to a small real-world service while keeping every run reproducible.

Their shared path list lives in src/extra-routes.mjs and every handler returns ok.

1. Ping
   • Request to [GET] /
   • Return Hi
   • Headers must contains text Content-Type: text/plain, additional context is acceptable eg. Content-Type: text/plain; charset=utf-8
2. Query
   • Request to [GET] /id/:id
   • Extract path parameter, query string and setting headers.
   • For this benchmark, the request URL will be send as: /id/1?name=bun
   • Headers must contains x-powered-by to benchmark
   • Expected response: "1 bun" (${id} ${query})
     • You MUST NOT use hardcode string or index to extract querystring.
     • In a real-world situation, there's no enforcement that the request will follow the specification, using hardcode index to extract name=bun querystring will be prone to error.
     • To test if it pass the requirement, the implementation should be able to extract querystring dynamically (please treat the value of 'name=bun' can be any value beside 'bun', for example 'alice', 'hina'), which means that the same code should be able to extract querystring, for example:
     • /id/1?name=bun&id=1 -> should return 1 bun not 1 bun&id=1
     • /id/1?id=1 -> should return 1
     • Query beside name maybe not need to be extracted and is optional
   • Headers must contains text Content-Type: text/plain, additional context is acceptable eg. Content-Type: text/plain; charset=utf-8
3. Body
   • [POST] /json
   • Mirror body to response
   • Server MUST parse body to JSON and serialize back to string
   • For the benchmark, the request body will be sent as: { "hello": "world" }
   • Expected response: { "hello": "world" }
   • Headers must contains text Content-Type: application/json, additional context is acceptable eg. Content-Type: application/json; charset=utf-8.
4. Video
   • [GET] /video
   • Stream public/kyuukurarin.mp4 without buffering the whole file in application memory.
   • Headers must contain Content-Type: video/mp4.
   • Uses 10 concurrent connections instead of 500 because the file is 14.1 MB.
   • Sends Cache-Control: no-store and a deliberately non-matching If-None-Match value, and requires 200, so every request transfers the full file even when the server emits an ETag.

Requirement

• The framework must at-least has latest published in less than 9 month otherwise will be classified as unmaintained and removed unless is an industry standard (Express).

Test machine specification

$ uname -a
Linux seia 7.0.11-1-cachyos #1 SMP PREEMPT_DYNAMIC Wed, 03 Jun 2026 22:05:15 +0000 x86_64 GNU/Linux

Results

These results are measured in req/s:

Framework	Runtime	Average	Ping	Query	Body	Video	Bundle Size	Startup	Memory Before/After
uws	node	267,311.582	427,899.11	409,892.79	231,035.02	419.41	2.3 KB	20.0 ms	70.8 / 139.9 MB
elysia-aot	bun	214,518.67	409,673.06	234,387.61	212,373.44	1,640.57	136.3 KB	16.0 ms	36.3 / 48.7 MB
elysia	bun	212,598.372	408,136.5	226,959.6	213,631.8	1,665.59	186.8 KB	19.6 ms	37.3 / 48.4 MB
h3	bun	172,978.673	268,455.21	219,564.22	202,299.58	1,595.68	31.1 KB	7.3 ms	33.6 / 50.0 MB
bun	bun	172,879.358	232,912.2	245,937.47	210,980.87	1,686.89	2.3 KB	4.0 ms	26.7 / 38.4 MB
hono	bun	169,752.09	271,679.6	221,518.09	184,179.03	1,631.64	19.7 KB	7.4 ms	33.5 / 44.3 MB
ultimate-express	node	168,817.505	424,303.7	118,309.22	132,197.69	459.41	583.5 KB	54.1 ms	107.0 / 240.8 MB
hono	deno	168,727.828	273,089.68	192,674.87	208,728.09	418.67	n/a	41.6 ms	66.2 / 98.6 MB
deno	deno	167,972.283	237,237.03	202,167.34	232,064.26	420.5	n/a	13.2 ms	58.5 / 89.2 MB
bun-web-standard	bun	162,686.943	231,235.21	199,156.8	218,736.32	1,619.44	1.7 KB	3.8 ms	27.1 / 41.7 MB
deno-web-standard	deno	161,145.583	232,952.03	176,955.08	234,253.33	421.89	n/a	14.4 ms	58.8 / 86.4 MB
hyper-express	node	138,848.412	226,731.93	181,959.05	146,334.69	367.98	247.2 KB	39.2 ms	77.9 / 207.3 MB
h3	deno	129,564.015	200,966.71	149,816.45	167,046.48	426.42	n/a	29.3 ms	73.8 / 163.0 MB
effect	bun	123,353.122	201,211.96	164,772.49	125,838.59	1,589.45	221.8 KB	32.5 ms	48.8 / 73.6 MB
fastify	bun	122,655.538	187,286.97	171,025.02	132,017.52	292.64	556.8 KB	76.8 ms	74.9 / 193.8 MB
fastify	node	103,292.998	155,273.66	143,072.85	114,475.25	350.23	556.8 KB	69.7 ms	95.7 / 205.0 MB
h3	node	98,758.288	149,801.48	133,825.29	111,055.35	351.03	50.4 KB	45.5 ms	85.3 / 190.1 MB
hono	node	93,107.478	146,046.63	121,489.54	104,659.13	234.61	52.8 KB	48.9 ms	97.0 / 220.6 MB
elysia-aot	node	92,923.678	144,966.13	120,779.71	105,726.96	221.91	153.4 KB	63.0 ms	91.5 / 182.8 MB
elysia	node	92,489.338	143,690.09	119,227.17	106,817.31	222.78	212.2 KB	58.6 ms	93.0 / 181.3 MB
effect	node	88,635.68	144,825.52	119,671.74	89,680.71	364.75	281.7 KB	53.5 ms	93.0 / 177.6 MB
express	bun	63,217.945	95,984.74	89,301.15	67,289.15	296.74	836.2 KB	42.5 ms	64.9 / 152.4 MB
koa	node	44,591.625	68,962.39	62,198.63	46,871.56	333.92	729.6 KB	51.7 ms	96.1 / 192.2 MB
express	node	42,293.953	64,589.19	57,320.85	46,922.84	342.93	616.3 KB	44.2 ms	87.2 / 208.8 MB
adonis	node	37,642.445	52,057.13	47,217.1	50,969.25	326.3	1.2 MB	351.8 ms	124.4 / 167.8 MB
nest	node	31,939.89	49,322.92	43,493.67	34,609.05	333.92	1.2 MB	116.5 ms	118.7 / 224.3 MB
oak	deno	24,449.013	35,570.6	35,004.12	26,977.52	243.81	n/a	60.2 ms	103.3 / 195.1 MB
acorn	deno	10,078.76	13,000.19	13,536.68	13,512.71	265.46	n/a	77.2 ms	106.4 / 237.5 MB

Note

1. uws, hyperexpress and ultimate-express bundle size is not accurate because uwebsocket is a native binary that can't be compiled to single bundle, and bundle size is vary based on operating system and CPU architecture
2. uws is a C++ framework with JavaScript binding

See more detail in results

Notice

I highly recommended testing this benchmark on your machine yourself as performance in likely to vary between machine.

If you are unable to run Deno, please run each Deno app individually first until the Deno finish installing the package, then proceed to run benchmark using bench.sh or npm run benchmark