Let’s run Julia everywhere from mobile to web

Satoshi Terasaki@AtelierArith

Link to slide ↓↓↓

Goma-chan (my family)

Goma-chan (projected into 2D space)

What I created

  • SubsetJuliaVM — a stack-based VM for a subset of Julia
    • Looks like Julia and walks like Python
    • CLI binary name: sjulia (think “subset Julia”)
    • Implemented in Rust
  • Runs on iOS and Android devices
    • The VM compiles to WebAssembly, so it also runs in the browser
  • Supports parts of Plots.jl, OrdinaryDiffEq.jl, Symbolics.jl, AbstractAlgebra.jl, and more

What I created

Deliverables

iOS / iPadOS — a native app with no JIT, fully App Store compliant

iOS app
App Store

Source code
julia-vm-oss

Web playground
subset_julia

Google it SubsetJuliaVM

Live demo (inside these slides)

  • Seeing is believing
  • The next slides embed the very same WASM VM.

Edit the code and run it with Ctrl+Enter / ⌘+Enter (or Run). Press Clear to clear the output.

Hello World

Fractals (Interact + Distributions + Plots)

Why I created it

  • I loved Pythonista (iOS app)
  • I wanted to run Julia on my iPad — offline
  • I was asked by some of the teachers who teach mathematics in Japan, “Can I use Julia on an iPad?”

What about Binder or Google Colab?

  • Unusable where the internet connection is unreliable
  • Installing external packages takes a long time

Can we run official Julia on iOS? (1/3)

  • It’s difficult to run the original functionality as-is.
    • Official Julia relies on JIT compilation
    • Ordinary iOS apps cannot freely generate and execute native code at runtime (W^X)
    • App Store rules also restrict downloaded code that changes app functionality (Review Guidelines §2.5.2)
  • SubsetJuliaVM uses a bundled interpreter instead: bytecode stays data as Vec<Instruction>.

Can we run official Julia on iOS? (2/3)

Experimentally, yes — but this is not an official iOS port.

Verified with Julia 1.12.6 on a physical iPhone 15 Pro Max:

println("Hello World")
using InteractiveUtils
versioninfo()
A = rand(2, 2)
x = rand(2)
y = A * x # <- does not work
  • Bundle the official macOS arm64 runtime and sysimage
  • Rewrite Mach-O platform metadata and sign every dylib
  • Run with --compile=no and small compatibility shims

Major limitations

  • Most standard libraries, packages, and JLLs are unsupported
  • BLAS-backed matrix operations still fail with `ccall` requires the compiler
  • Development-signed research prototype; not App Store-ready

Can we run official Julia on iOS? (3/3)

How I created SubsetJuliaVM

  • Apps like Pythonista and Carnets do run on iOS.
    • So why not build a virtual machine that accepts Julia syntax?!
  • I don’t have the ability to create this 😭
    • Let’s use AI 🤖
    • Built with heavy use of AI tools: Claude Code, Codex, Cursor, DeepSeek, Kimi, etc.

Let’s actually build SubsetJuliaVM!

Julia source
→
parse → CST
→
lower → Core IR
→
compile → bytecode
→
stack VM → result

Result: AtelierArith/julia-vm-oss

$ git clone https://github.com/AtelierArith/julia-vm-oss.git
$ cd julia-vm-oss
$ ./scripts/sjulia_install.sh   # Windows: pwsh -File scripts/sjulia_install.ps1
$ sjulia examples/mandelbrot.jl
  0.027875 seconds
Mandelbrot Set (50x25):

                              .
                              . .
                              .+
                             ###+.
                        .   .####.
                       .#++#########....
                      ..##############.
            .        ..################..
             ...... .##################.
            .#######.###################
          ...##########################.
#####################################..
          ...##########################.
            .#######.###################
             ...... .##################.
            .        ..################..
                      ..##############.
                       .#++#########....
                        .   .####.
                             ###+.
                              .+
                              . .
                              .

How SubsetJuliaVM differs from official Julia

How code is interpreted and executed:

Each cell: who does it → what it produces

Stage Official Julia SubsetJuliaVM
Parsing flisp / JuliaSyntax.jl → AST pure-Rust parser → CST
Lowering Julia’s lowering → lowered IR Rust lowering → Core IR
Code generation LLVM JIT at runtime → native code bytecode compiler before execution → bytecode
Execution CPU runs native code stack VM interprets bytecode
Runtime library C runtime + pure-Julia base/ Rust builtins + a pure-Julia subset
  • With no JIT, SubsetJuliaVM runs even on iOS (W^X) and WASM

What SubsetJuliaVM does not provide

Everything that would break the sandboxed, offline, no-downloaded-code model:

  • Pkg.jl / package management — no Pkg.add, no registry, no downloads; using X only reaches the ~34 packages bundled into the binary
  • Network I/O — no Sockets, no Downloads/HTTP
  • Filesystem access — file operations are unimplemented; include resolves through virtual embedded paths
  • External processes — no run(cmd)
  • C interface — no @ccall/pointer/unsafe_*

What we support

  • Basic Julia syntax (subset)
  • ~34 bundled 3rd-party packages (no download), including:
    • Plots, StatsPlots, JSXGraph, Interact
    • OrdinaryDiffEq, SciMLBase
    • Symbolics, AbstractAlgebra, Primes, Combinatorics
    • Distributions, StatsBase, Optim, QuadGK, HCubature
    • StaticArrays, SparseArrays, SpecialFunctions, Quaternions, Rotations, …
  • These have been re-implemented specifically for SubsetJuliaVM.

What we support

juliars

An experimental Julia → Rust AoT transpiler (same frontend as the VM).

  • Pipeline: source → CST → Core IR → Rust source → rustc -O → native binary
  • Unlike JuliaC / PackageCompiler, it does not need libjulia
  • Best for scalar numerics, loops, arrays — the broad package surface stays on the VM
  • On the coprime-π / Mandelbrot benches: about as fast as official Julia
$ cargo build --release -p subset_julia_vm --features aot --bin juliars
$ target/release/juliars --minimal-prelude examples/mandelbrot.jl --emit-binary target/mandelbrot_aot
$ target/mandelbrot_aot

Performance: Julia vs. SubsetJuliaVM

Coprime-π benchmark — \(P(\gcd(a,b)=1)=6/\pi^2\) (Apple Silicon macOS):

# P(gcd(a,b)=1) = 6/π²  →  π = √(6/P)
function mygcd(a, b)
    while b != 0
        tmp = b
        b = a % b
        a = tmp
    end
    a
end

function calc_pi(N)
    cnt = 0
    for a in 1:N
        for b in 1:N
            if mygcd(a, b) == 1
                cnt += 1
            end
        end
    end
    sqrt(6.0 / (cnt / N / N))
end

calc_pi(10000)  # same source for all runtimes
Runtime Time
Official Julia 2.12 s
juliars (AoT) 1.89 s
sjulia (VM) 9.99 s
Python 3.14 16.76 s
  • AoT is about as fast as official Julia

Interested in the Android app?

  • Coming soon — we are looking for testers.
  • Please contact s.terasaki@atelier-arith.jp

Conclusion

Thank you!!!

Q&A material: see the Appendix slides