Jarred Sumner
@jarredsumner

building @bunjavascript at @anthropicai. formerly: @stripe (twice) @thielfellowship. high school dropout. npm i -g bun

 San Francisco, CA
bun.com
Joined May 2014
3,979 Photos and videos
People frequently ask: > How is Bun sustainable? If I bet my company’s tech stack on Bun, will Bun still be around in a few years? We didn’t have a great answer to this question, until today
Bun
@bunjavascript
2 Dec 2025
Bun is joining Anthropic! bun.com/blog/bun-joins-anthr…
266
143
4,733
1,356,380
Jarred Sumner retweeted
Bun
@bunjavascript
Oct 1
Today, Bun is a JavaScript runtime, bundler, test runner, and a package manager. Soon, Bun will also be a
201
29
957
77,920
An early preview of experimental Bun + JSC with ahead-of-time compilation For Claude Code's CLI: 24% faster startup, 48% lower memory, 46% lower CPU, and nearly 2x larger binary size
Bun AoT VS Bun JIT Claude Code, same source, same 5-turn session. AoT runs with the JIT off. Time to interactive 165 ms 216 ms -24% Wall time launch + 5 turns 322 ms 529 ms -39% CPU time Memory at interactive 395 ms 738 ms -46% 56 MB 108 MB -48% Memory after 5 turns 82 MB 159 MB -48% 101 MB Peak memory Executable size target: +100 MB at most 194 MB -48% 430 MB 229 MB +88% Early prototype, macoS arm64. Medians of 8 runs against a fake API server on one MacBook Pro. Memory is phys_footprint. CPU is user + system, all threads. Turns in the JIT build are not CPU bound, so part of the wall time gap is waiting. 28 Sep 2026.

ALT Bun AoT VS Bun JIT Claude Code, same source, same 5-turn session. AoT runs with the JIT off. Time to interactive 165 ms 216 ms -24% Wall time launch + 5 turns 322 ms 529 ms -39% CPU time Memory at interactive 395 ms 738 ms -46% 56 MB 108 MB -48% Memory after 5 turns 82 MB 159 MB -48% 101 MB Peak memory Executable size target: +100 MB at most 194 MB -48% 430 MB 229 MB +88% Early prototype, macoS arm64. Medians of 8 runs against a fake API server on one MacBook Pro. Memory is phys_footprint. CPU is user + system, all threads. Turns in the JIT build are not CPU bound, so part of the wall time gap is waiting. 28 Sep 2026.

69
62
1,300
84,512
more replies
There are a lot of optimizations that aren’t implemented yet. I want the binary size growth to go down a lot more (it started at 1 GB). It needs to support sized types like uint32_t and propagate more detailed types. There’s no separate checker to run and no diagnostics
3
50
6,103
AoT uses Bun’s bundler & parser to parse typescript, tags types to JavaScriptCore via a typecheck bytecode intrinsic, and compiles the bytecode to machine code via B3 (what JSC’s JITs use) This keeps the JavaScriptCore changes well-contained, so that syncing with upstream WebKit isn’t complicated
1
56
5,672
Garbage collector scheduling adjustments make Bun v1.4.3 use significantly less idle CPU, slightly reduce memory, and are performance ~neutral at throughput
Bun
@bunjavascript
Sep 27
In the next version of Bun Idle CPU usage drops by 67%
Bun v1.4.3: 67–89% less CPU while your server sleeps, compared with v1.4.2. Illustrated as a night sky with the Bun logo asleep as the moon. One idle minute: v1.4.2 wakes up every second, because any occasional request kept the 1-second GC timer running. v1.4.3 wakes only for the health check, every 10 seconds. CPU milliseconds used in 2 idle minutes on Bun 1.3.14, 1.4.0, 1.4.2 and 1.4.3: Express: 430, 308, 282, 57 (4.9× less than 1.4.2) Fastify: 450, 284, 267, 57 (4.7× less) Elysia: 473, 280, 197, 21 (9.4× less) Hono: 417, 263, 195, 21 (9.3× less) Next.js SSR: 7,029, 628, 815, 268 Also 6–11% less idle memory than v1.4.2. Measured after a 30-second load test with one health check every 10 seconds, on Linux x64. 1.3.14 and 1.4.0 are the mean of 2 runs; 1.4.2 and 1.4.3 are the median of 4.

ALT Bun v1.4.3: 67–89% less CPU while your server sleeps, compared with v1.4.2. Illustrated as a night sky with the Bun logo asleep as the moon. One idle minute: v1.4.2 wakes up every second, because any occasional request kept the 1-second GC timer running. v1.4.3 wakes only for the health check, every 10 seconds. CPU milliseconds used in 2 idle minutes on Bun 1.3.14, 1.4.0, 1.4.2 and 1.4.3: Express: 430, 308, 282, 57 (4.9× less than 1.4.2) Fastify: 450, 284, 267, 57 (4.7× less) Elysia: 473, 280, 197, 21 (9.4× less) Hono: 417, 263, 195, 21 (9.3× less) Next.js SSR: 7,029, 628, 815, 268 Also 6–11% less idle memory than v1.4.2. Measured after a 30-second load test with one health check every 10 seconds, on Linux x64. 1.3.14 and 1.4.0 are the mean of 2 runs; 1.4.2 and 1.4.3 are the median of 4.

14
11
667
34,943
sounds like a Bun 2.0
Jarred Sumner
@jarredsumner
Sep 24
JavaScript was never designed for the server. If that doesn't change soon, JavaScript on the server will be replaced by Rust. The fix needs to happen in the engine. Sound types. Ahead-of-time compilation. Threads with shared objects.
70
21
1,344
140,385
to be clear, if we do add this we would only add it in a backwards compatible way. just like the rest of bun, compatible with the existing ecosystem from the start.
10
183
11,853
JavaScript was never designed for the server. If that doesn't change soon, JavaScript on the server will be replaced by Rust. The fix needs to happen in the engine. Sound types. Ahead-of-time compilation. Threads with shared objects.
Arto Bendiken
@bendiken
Sep 23
"I love Rust! Rust is amazing ...if you never, ever, EVER have to look at it yourself." — @dhh
303
163
3,850
906,850
JavaScript has been through this before. In 2011, Google set out to replace JavaScript with Dart, arguing JavaScript couldn't be fixed by evolving it. Then ES2015 shipped. It's time to do that again.
13
8
575
46,650
JavaScript engines like JavaScriptCore and V8 have some of the best compiler tech in the world. An open question: with sound types, ahead-of-time compilation for statically typed code, and JITs for everything else, can JavaScript be competitive with Rust and C?
60
5
711
40,280
Opus 5.5 is a very good model and I expect it will quickly become the daily driver for a lot of people. The writing style is so much better. It writes like Opus 4.6 and codes like Fable 5.1.
Claude
@claudeai
Sep 22
Introducing Claude Opus 5.5, the first model in our new Claude 5.5 family. It performs at the level of Claude Fable 5.1 for most tasks, and costs 40% less to run than Opus 5.
39
34
1,487
59,512
Bun v1.4.3 ships on Monday
27
5
386
20,990
In the next version of Bun `Bun.FetchSession` gives you a `fetch` function with its own keepalive connection pool
Dark slide with the Bun logo and a "new in Bun" badge. Large headline: "Bun.FetchSession". Subtitle: "fetch() with its own connection pool." Below it, a code sample: const api = new Bun.FetchSession({ proxy: "http://proxy:8080", tls: { ca }, keepAlive: { idleTimeout: 30 }, }); await api.fetch(url); new Octokit({ request: { fetch: api.fetch }, }); "api.fetch" is highlighted in both places. Three cards at the bottom read: "Own pool: isolated keep-alive", "Proxy + TLS: set once per session", "Drop-in: works in any SDK".

ALT Dark slide with the Bun logo and a "new in Bun" badge. Large headline: "Bun.FetchSession". Subtitle: "fetch() with its own connection pool." Below it, a code sample: const api = new Bun.FetchSession({ proxy: "http://proxy:8080", tls: { ca }, keepAlive: { idleTimeout: 30 }, }); await api.fetch(url); new Octokit({ request: { fetch: api.fetch }, }); "api.fetch" is highlighted in both places. Three cards at the bottom read: "Own pool: isolated keep-alive", "Proxy + TLS: set once per session", "Drop-in: works in any SDK".

23
19
628
45,942
PHP got one thing right JavaScript never had In next week’s version of Bun
84
35
1,500
213,867
Claude Code's idle memory dropped 14% on Linux with Bun v1.4.1
Line chart: "Claude Code — p90 idle memory usage (Linux)", headline "−14%, lower is better". Two lines plot 90th-percentile memory of idle Claude Code processes against process uptime, from about 1 hour to 2 days. The grey "Before" line rises from 490 MiB to 665 MiB; the orange "Bun v1.4.1" line stays below it the whole way, rising from 431 MiB to 572 MiB. Footnote: p90 RSS after 3+ min idle, Linux, Claude Code 2.1.260+ vs earlier versions, same days and session age, Sep 4–9, 2026; y-axis starts at 250 MiB.

ALT Line chart: "Claude Code — p90 idle memory usage (Linux)", headline "−14%, lower is better". Two lines plot 90th-percentile memory of idle Claude Code processes against process uptime, from about 1 hour to 2 days. The grey "Before" line rises from 490 MiB to 665 MiB; the orange "Bun v1.4.1" line stays below it the whole way, rising from 431 MiB to 572 MiB. Footnote: p90 RSS after 3+ min idle, Linux, Claude Code 2.1.260+ vs earlier versions, same days and session age, Sep 4–9, 2026; y-axis starts at 250 MiB.

23
5
523
23,264
Excited to see everyone tomorrow! We’re at capacity with 130+ on the waitlist, so if you can’t make it please mark “can’t go” on Partiful so someone else can grab your spot
Jarred Sumner
@jarredsumner
Aug 28
We're throwing a party in SF for Bun v1.4 on September 9th. If you use Bun (or want to), come hang out partiful.com/e/GJtQ5hI0PaNDy…
3
2
129
22,366
Load more