Skip to main content
Classiq hosts public logical noise models that any user can reference by name, with no simulation to run yourself. This is the fastest way to get started — retrieve one with get_logical_noise:
More public models can be added by request, and will be introduced over time. To model hardware that none of them describes, build your own instead — see Custom Noise Models.

​
The models

Λ is the error-suppression factor per two units of code distance — how much the logical error rate improves each time you grow the code. Larger is better, and Λ ≤ 1 means the code no longer suppresses errors at all.
The two device-flavored models are illustrative reference points, not vendor calibration snapshots. They are rounded from published figures to give you a feel for how real superconducting hardware behaves; they are not tracked against any vendor’s live calibration data, and no vendor publishes every parameter the engine needs.

​
Physical parameters

See Custom Noise Models for what each field means and which rules it accepts.

​
public/depolarize_1e-3

A general-purpose depolarizing model, generated from custom_noise_model(1e-3) as defined in Custom Noise Models. Use it as a well-behaved baseline rather than as a model of any particular machine.

​
public/google_willow_like

Flavored after a superconducting processor tuned for error correction: fast surface-code cycle, comparatively short coherence, and a dedicated multi-level reset with leakage removal — so its reset error is well below its readout error.
The single-qubit, two-qubit and measurement values are rounded from Google’s published Willow spec sheet for its error-correction chip (0.035%, 0.33% CZ, and 0.77% measurement, all measured under simultaneous operation). The idle terms are derived from the published mean T1 of 68 µs against the gate and cycle durations; reset is an assumption.

​
public/ibm_heron_r2_like

Flavored after a superconducting processor with roughly three times the coherence but slower, measurement-limited readout and reset.
The two-qubit value uses a published layered error figure rather than an isolated median, because a surface code runs its two-qubit gates simultaneously and the isolated figure understates what the code actually sees. The idle terms are derived from a published median T1 on the order of 200 µs against the gate and readout durations; reset is assumed to be measurement-limited.
Both device-flavored models are deliberately conservative — their simulated Λ comes out below what the corresponding real hardware achieves. Their T gate error rates are also high enough that magic-state distillation is impractical at these noise levels, so treat them as a guide to Clifford-level behavior rather than as a basis for end-to-end resource estimates.

​
References

The device-flavored models are rounded from the public sources below. All of them are point-in-time figures — vendors recalibrate continuously, and these models are not tracked against live calibration data.