⚡️ tzap

Installation · Using tzap · Qiskit integration · PennyLane integration

A super fast, Rust-based optimizer for large Clifford+T circuits.

• tzap is state-of-the-art in speed, scalability, and gate-count reduction.
• tzap minimizes T-count with a new linear-time phase folding algorithm, based on this paper.
• tzap implements a new and fast superoptimization pass.
• The core optimization algorithms are fully formalized in Lean under formalization.

tzap is multiple orders of magnitude faster than other optimizers—and linearly scales to millions of gates! Here's a runtime comparison to two powerful optimizers on increasingly larger circuits.

Installation

You can use tzap as a command-line utility or a library.

Install the binary

These options install the standalone native tzap executable.

Homebrew (macOS/Linux):

brew install qqq-wisc/tap/tzap

Prebuilt release binary (macOS/Linux):

curl -LsSf https://github.com/qqq-wisc/tzap/releases/latest/download/tzap-opt-installer.sh | sh

You can also build and install tzap from crates.io (cargo install tzap-opt) or build from source (cargo install --path .). You can also use tzap through the Rust API; see the Rust API documentation.

Integrations with Qiskit and PennyLane

You can also use tzap as a Python library and apply it as Qiskit optimization pass or PennyLane transform. See the Qiskit API guide or PennyLane API guide for framework-specific setup.

Running tzap

The standard command-line workflow is described below.

Optimize a circuit

tzap input.qasm -o output.qasm

For example, using a benchmark in this repo:

$ tzap benchmarks/feynman/hwb12.qasm -o optimized.qasm
⚡️ tzap v0.4.3
  Parsed benchmarks/feynman/hwb12.qasm (5.5 MB) in 0.098s
	└─ 20 qubits · 514,412 gates
  Loaded superoptimizer table in 0.018s

  Converged after 6 rounds

  ┌─ Final result · 43.7% fewer gates · 2.169s ──────────────────────────┐
  │ Gates    ━━━━━━━━━━━━━╸────────────────── ↓43.7% · 514,412 → 289,484 │
  │ 2q gates ━━━━━╸────────────────────────── ↓18.7% · 191,803 → 155,914 │
  │ T/Tdg    ━━━━━━━━━━━━━━━╸──────────────── ↓49.9% · 171,465 →  85,897 │
  │ Depth    ━━━━━━━╸──────────────────────── ↓24.3% · 274,781 → 207,940 │
  └──────────────────────────────────────────────────────────────────────┘
  wrote optimized.qasm

Optimization levels

tzap benchmarks/feynman/hwb12.qasm -O1 -o optimized.qasm

Python APIs

The package exposes the native optimizer through optimize_qasm, a Qiskit optimization pass, and a PennyLane transform.

Decompose Rz into Clifford+T

Use --decompose-rz when the target backend only accepts Clifford+T; tzap uses gridsynth. --epsilon trades approximation accuracy for circuit size (default 1e-10; larger is coarser).

tzap input.qasm -o output.qasm --decompose-rz --epsilon 1e-6

Use --decompose-cz to decompose CZ gates into H+CX+H before the optimization pipeline.

Custom pipeline

--passes runs an explicit, ordered sequence of passes in place of the default pipeline.

tzap input.qasm -o output.qasm --passes CancelGates,PhaseFoldRand
tzap input.qasm -o output.qasm --passes DecomposeCz,CancelGates,PhaseFoldRand

Circuit support

tzap supports a subset of OpenQASM 2.0:

• Gates: h, x, z, s, sdg, t, tdg, rz, cx, ccx, ccz, cz, measure, reset
• Declarations: qreg, creg
• Not supported: classical conditionals (if), custom gate definitions (gate), barriers, include files (besides qelib1.inc, which is ignored)
• Unrecognized lines produce an error

Toffoli (ccx) and doubly controlled-Z (ccz) are auto-decomposed into Clifford+T. Controlled-Z (cz) is kept native so phase folding and cancellation can operate through it; use --decompose-cz for H+CX output. Rz is left as-is unless you pass --decompose-rz.

Correctness

1. Fuzzing and equivalence verification on small random circuits and benchmark circuits.
2. Lean formalization: core algorithms are implemented and proven sound in Lean 4 — see formalization.

Citation

If you use tzap in your research, please cite:

@misc{albarghouthi2026tzap,
      title={Linear-Time T-Gate Optimization via Random Abstraction}, 
      author={Aws Albarghouthi},
      year={2026},
      eprint={2605.13929},
      archivePrefix={arXiv},
      primaryClass={cs.PL},
      url={https://arxiv.org/abs/2605.13929}, 
}