The compiler that deletes the runtime.

ZeptonML lowers an exported PyTorch graph to a self-contained native binary. Memory is planned at compile time; the output links nothing it does not use.

01
PyTorch modeltrained, in memory
02
torch.exporttraced graph
03
Computational graphops + tensors
04
ZeptonML · Zigplan + codegen
05
Static binarydistroless OCI
Inputtorch.export ExportedProgram (.pt2)
BackendZig — no libc dependency required
Targetsx86-64 · aarch64 · riscv64 · bare metal
OutputStatic binary or distroless OCI image
MemoryArena planned at compile time · zero GC
Runtime depsNone
01

Whole-graph planning

Allocation, layout, and op fusion are decided once, ahead of time — not rediscovered every inference.

02

Dead weight removed

Only the operators your graph actually calls are emitted. Nothing else is linked.