Free quantum computer simulators you can use today

Run a circuit now

The example below is a Bell state: two qubits prepared in one shared state, then measured 1,000 times. Each repetition is a shot. On an ideal simulator about half the shots read 00 and the rest read 11.

# 'backend' is pre-created for you in the Qollab Playground.
from qiskit import QuantumCircuit

qc = QuantumCircuit(2, 2)       # two qubits, two answer bits
qc.h(0)                         # H gate on qubit 0
qc.cx(0, 1)                     # CNOT: qubit 1 follows qubit 0
qc.measure([0, 1], [0, 1])      # read both qubits into the bits

job = backend.run(qc, shots=1000)
counts = job.result().get_counts()
print(counts)                   # e.g. {'00': 502, '11': 498}

To run it:

1. Press Run on Qollab under the code. The project page opens in a new tab.
2. Press Run on Qollab on that page. The Run dialog opens, headed Select QPU.
3. Pick Built-in simulator and press Run. The counts print in the console.

Without an account you get one run, on one simulator. Sign in with Google or GitHub to run again and to try the other simulators. The playground runs Python through a browser feature called WebAssembly JSPI, found in a recent Chrome, Edge or Opera, or Firefox 153 or later. In other browsers Qollab keeps Run disabled. The first run loads Python and the example's packages, so it takes longer than the runs after it.

The same circuit in JavaScript runs the same way. Its main.js calls the same Qiskit from a JS project.

The simulators in the Run dialog

All 14 are free. Signed out, the dialog marks the IonQ rows Credits apply; the runs still cost nothing. The IBM figures are the sizes of the devices the models copy; how large a circuit your browser can simulate depends on its memory. IonQ has retired the Aria hardware, so the two Aria models copy machines that no longer run. When IonQ reports its machines, the dialog also lists its Forte quantum computers, which cost credits. Compute backends lists every option.

What the counts show

The Bell state above, run on Qollab on 15 October 2026, signed out, 1,000 shots each:

Every run but one took 5 to 18 seconds from pressing Run to the counts, in a fresh browser. One IonQ cloud run took 133 seconds.

The ideal simulators return only 00 and 11. Their split differs from 500/500 by chance, as any 1,000 samples of a 50/50 outcome do. The noise-model simulators also return 01 and 10, outcomes the ideal circuit never produces. Across the 12 models those two outcomes took between 0.4% and 30.3% of the shots, because each model copies a different device's calibration.

Each noise model adds errors measured on its device. IBM builds its models from calibration snapshots that IBM says were sometimes taken years ago. IonQ calls its models simplified approximations and says hardware results should not be expected to match them quantitatively. One run of one circuit on these models does not rank the devices or their makers. Noise-model runs usually take longer than ideal ones; Ideal vs. noise-model simulators explains when and why.

Which simulator fits which job

• Learning what gates do: a drag-and-drop simulator in the browser, such as Quirk or IBM Quantum Composer, updates the state as you place each gate.
• Checking a circuit's logic: an ideal simulator, such as the built-in one on Qollab, or Qiskit, Cirq or PennyLane on your machine.
• Seeing noise before a hardware run: a noise-model simulator. On Qollab that is any IBM or IonQ model; elsewhere, IBM's fake backends with Qiskit Aer, IonQ's cloud simulator, or the density-matrix simulators in Cirq, PennyLane and Braket.
• Going past 30 qubits: a GPU simulator such as CUDA-Q or PennyLane Lightning, a tensor-network simulator for circuits with limited entanglement, or a cloud service.
• Error-correction research: Stim, which simulates Clifford circuits with thousands of qubits.

Free quantum simulators compared

Each tool below was checked against its maintainer's own pages on 14 October 2026. Free tiers change, so check a tool's pricing page before you plan around it. A hard qubit limit means the tool refuses more; a practical one means memory sets the limit.

In a browser tab

On your machine

On a provider's cloud

Specialist simulators

• Stim simulates Clifford circuits with Pauli noise, at thousands of qubits and millions of operations. Error-correction research uses it.
• QuEST (C and C++) and Qulacs (Python and C++) simulate state vectors and density matrices, with GPU support.
• MQT DDSIM, from the Technical University of Munich, simulates circuits with decision diagrams and plugs into Qiskit.
• QuTiP models open quantum systems; its qutip-qip package adds gate circuits and pulse-level noisy devices.
• Quantum Inspire 2.0, from QuTech, runs a cloud emulator free up to 26 qubits for non-commercial users, after account approval.

Paid managed simulators

Amazon Braket's managed simulators, SV1 (state vector, up to 34 qubits) and DM1 (density matrix, up to 17), cost $0.075 a minute in US East (N. Virginia). Billing is per millisecond, with a three-second minimum per simulation. AWS changed its Free Tier for new accounts in July 2025, adding sign-up credits, and Braket's pricing page still describes a free simulator hour. Check what your account gets before a long run.

Retired or changed since 2024

• IBM retired its cloud simulators on 15 May 2024. It points users to the fake backends and Qiskit Aer on their own machines.
• Azure Quantum Credits ended on 1 June 2025.
• Xanadu's Quantum Cloud is no longer available; the decommissioning was committed in January 2026, and the Strawberry Fields repository is archived.
• Amazon Braket retired its TN1 tensor-network simulator in June 2026.
• Classiq ended open free access for new users on 9 August 2026; existing users keep access.
• Microsoft removed the Q# katas from its online playground in September 2026. They now run in the QDK extension for VS Code.

How many qubits can a simulator handle?

A state-vector simulator stores 2^n complex numbers for n qubits, at 16 bytes each in double precision. Each added qubit doubles the memory:

A density matrix, which noise-model simulators can use, holds 4^n numbers: 15 qubits take the same 16 GiB as a 30-qubit state vector. Qollab's built-in simulator stops at 24 qubits, the state-vector limit of Qiskit's BasicSimulator. AWS gives 25 qubits as practical guidance for its local simulator.

Other methods trade generality for size. Stim is a stabilizer simulator: it handles thousands of qubits, but only for Clifford circuits (H, S, CNOT and measurement, with no T gate). The QDK's stabilizer simulator also accepts a small number of non-Clifford operations.

Tensor-network and matrix-product-state simulators reach hundreds or thousands of qubits on circuits whose entanglement stays limited. GPU simulators keep the state vector in GPU memory; NVIDIA's guidance for CUDA-Q is under 33 qubits on one 64 GB GPU. IBM's documentation gives several reasons for retiring its cloud simulators, among them that their usefulness "maxes out at around 50 qubits".

Questions people ask

Related

• Compute backends
• Ideal vs. noise-model simulators
• Noise, and the real machine