Python's ease.
Systems speed.
Zyth compiles your program to machine code with LLVM and ships it with its runtime in one file. Here is how that compares to Go, Rust and Java on the same six programs, and where Zyth is still behind.
faster start than the JVM: 1.6 ms against 24.2.
less memory than Java: 5 to 26 MB against 38 to 82.
programs faster than Go, built for release.
Rust's start-up time on a Mac: 1.41 ms against 1.36.
Milliseconds per run, lower is better
Whole runs, best of 15, pinned to one core, Linux (WSL2), 2026-11-01. Zyth zyth build --release (LLVM, static runtime), Go 1.25, Rust 1.99 rustc -O, JDK 25.0.5 with 4 cores.
The numbers
| Program | What it does | Zyth | Rust | Go | Java |
|---|---|---|---|---|---|
| empty | start and exit | 1.6 | 1.7 | 1.9 | 24.2 |
| fib | recursive fib(30) | 3.2 | 3.4 | 5.3 | 27.4 |
| words | count 300,000 words in a map | 21.1 | 21.2 | 24.4 | 70 |
| sieve | primes in a 2M-element list | 5.7 | 5.8 | 7.4 | 40.3 |
| text | split and parse text lines | 25.3 | 18.7 | 37.8 | 126 |
| structs | update a list of structs | 1.8 | 1.9 | 3.6 | 55.2 |
| sort | sort 300,000 numbers | 18.7 | 10.7 | 71.9 | 110.5 |
Fastest in each row in green. Lower is better.
A web server against Go and Node
The same JSON service in Zyth (zyth.net, built with zyth build --release), Go's net/http and Node's http module. Each server pinned to 2 cores, the load generator on other cores, 64 connections, median of 5 runs of 10 seconds, no errors.
| Server | /json, requests/s | p99 | parse a posted order, requests/s | p99 |
|---|---|---|---|---|
| Zyth | 288,000 | 0.59 ms | 218,000 | 0.74 ms |
| Go net/http | 189,000 | 1.42 ms | 131,000 | 2.05 ms |
| Node, 4 worker processes | 138,000 | 1.2 ms | 93,000 | 1.75 ms |
| Node, one process | 28,700 | 4.1 ms | 19,400 | 6.1 ms |
JSON is read and written straight between text and Zyth values, and each connection is a task on Zyth's work-stealing pool. The standard library also has HTML templates escaped by default, cookies and signed sessions, SQLite built in, a PostgreSQL client with TLS, and migrations. A program that uses no database gets a runtime without them.
use zyth.net.{server}
use zyth.db.{sqlite}
let db = sqlite.open("notes.db")
db.migrate("migrations")
server(port: 8080) { app ->
app.get("/api/posts") { _ -> db.query("select id, title from posts") }
}
GPU drawing against three.js
The asteroid belt of the solar system demo on its own: every frame, each asteroid's position on its orbit is computed and drawn. Zyth computes the positions as GPU tensors and draws them from the same memory. three.js (0.187, WebGL 2) runs it two ways: the usual one, a JavaScript loop that writes the positions and uploads them, and the orbit written by hand into its vertex shader. Windows 11, RTX 3090, Chrome 151, 1280 × 720, frame-rate cap off, frames counted over 10 seconds.
| Asteroids | Zyth, GPU tensors | three.js, positions from JavaScript | three.js, orbit in its shader |
|---|---|---|---|
| 24,000 | 1,791 fps | 776 fps | 6,792 fps |
| 250,000 | 1,116 fps | 74 fps | 5,273 fps |
| 1,000,000 | 519 fps | 19 fps | 1,941 fps |
| 4,000,000 | 176 fps | 4.6 fps | 280 fps |
Zyth is 2.3 to 38 times faster than the usual three.js way, because nothing goes back through JavaScript. The hand-written shader is faster than both: it does the orbit math inside the draw, where Zyth runs it as a few small tensor kernels first. It only works because this motion has a formula that fits in a vertex shader; Zyth's version is ordinary tensor code that could be any computation, a physics step or a neural network.
Where Zyth is behind
Web servers. Zyth's server has no HTTP/2, no TLS of its own and no streaming bodies yet; put it behind a proxy such as Caddy or nginx. The numbers above are from one machine over loopback, on 2 cores; at 256 connections Zyth's lead over Go on posted orders drops from 1.67x to 1.29x.
Text and sorting against Rust. Rust splits text 1.35x faster and sorts 1.75x faster. Zyth's lists hold each number as a tagged value, so a sort moves four times Rust's memory. Storing lists of numbers as plain arrays is planned.
Java after warm-up. Java's JIT gets faster as a program runs. After 40 runs in the same JVM, Java counts words 3.3x faster than Zyth and splits text 2x faster. Short programs and tools favour Zyth; long-running servers have not been measured yet.
Deep call trees. On some recursive code Zyth is about 1.4x Rust. That gap is Zyth's guarantees, not its compiler: a C program with the same three checks runs within 3% of Zyth.
What Zyth checks that Rust's release build does not
- Overflow is an error, never a wrong number. Rust's release build wraps silently.
- Runaway recursion is an error at 3,000 calls, with the call chain, never a crashed process.
- Every failure is a value the caller sees, kept in a register so it costs almost nothing.
What ships
| Language | What you copy to a server | Peak memory |
|---|---|---|
| Zyth | one 9.9 MB file | 5 to 26 MB |
| Java | a JAR plus a 43 MB runtime, or an 8 to 18 MB native image | 38 to 82 MB |
Try the same programs
zyth build --release sort.zyth -o sort
./sort