Google Willow

Overview

Google Willow is the chip that turned a decades-old theoretical promise into an experimental result. Quantum error correction predicts that if physical qubits are good enough, then combining many of them into a logical qubit can drive the logical error rate down exponentially. Willow, announced in December 2024, was the first superconducting processor to show that behavior in practice, as the error-correcting code was scaled from 3 by 3 to 5 by 5 to 7 by 7 grids of data qubits.

The chip has 105 qubits, roughly double Sycamore's 53, and substantially better coherence. Google also used it for a benchmark, random circuit sampling, and in October 2025 for the Quantum Echoes algorithm, which the company described as the first verifiable quantum advantage.

History and background

Google Quantum AI laid out a six-milestone roadmap toward a large-scale error-corrected quantum computer. Sycamore achieved the first milestone, a beyond-classical computation, in 2019. The second milestone, a prototype error-corrected qubit, came in 2023 when a distance-5 surface code slightly outperformed a distance-3 code. Willow delivered the more convincing version of that milestone in December 2024, with error suppression across three code sizes.

Since then Willow has been used for further experiments. In October 2025 Google reported Quantum Echoes in Nature, an out-of-time-order correlator measurement that ran about 13,000 times faster than the best classical method Google tested. In March 2026 the company opened an early-access program for outside research proposals on Willow, though the hardware is still not on a public cloud.

How it works

Willow's qubits are transmons arranged on a square grid, with tunable couplers between neighbors. In an error-correction experiment, some qubits hold the data of the logical qubit and others, called measurement qubits, repeatedly check parity conditions between neighbors. These checks reveal where errors have occurred without revealing the data. A classical decoder reads the check results in real time, infers the likely errors, and updates its tracking so that logical information stays protected.

In the surface code, a distance-d code uses about d squared data qubits plus a similar number of measurement qubits. A larger code can correct more errors, but it also exposes more qubits to noise. If the physical error rate is below a threshold, the larger code wins. Willow measured logical error per cycle at distances 3, 5, and 7 and found that each step up cut the error by roughly half, with the d = 7 memory lasting more than twice as long as its best physical qubit.

Google reports different configurations for different experiments. For error correction, the mean T1 was about 68 microseconds and the two-qubit gate error about 0.33%. For random circuit sampling, a configuration tuned for the benchmark reached about 98 microseconds mean T1 and 0.14% two-qubit error. The qubit grid has 105 sites, and the sampling experiment used 103 of them.

Connectivity

2D grid: each qubit connects to up to four nearest neighbors (average connectivity of about 3.47)

Specifications in detail

Key features

• Below-threshold surface-code error correction at distances 3, 5, and 7
• Mean coherence time about five times better than Sycamore
• Random circuit sampling on 103 qubits at depth 40 in under five minutes
• Used for the 2025 'Quantum Echoes' experiment, reported as a verifiable beyond-classical result
• Fabricated in Google's dedicated facility in Santa Barbara, California

Limitations

• Below-threshold results so far apply to quantum memory, not to logical gates
• Logical error rates are still far above the roughly one in a million needed for large algorithms
• Not publicly available; access has been limited to Google teams and selected research programs
• Random circuit sampling has no known practical use

Timeline

What it is used for

• Quantum error-correction research, especially surface-code memory and, next, logical gates
• Beyond-classical benchmarks such as random circuit sampling and out-of-time-order correlators
• Physics simulations that probe quantum dynamics, with proof-of-principle links to molecular NMR
• Developing the real-time decoders needed for fault-tolerant machines

How it compares

• Compared with IBM Heron, Willow has fewer qubits but better coherence and a design focused on surface-code memory, while Heron is optimized for broad cloud access.
• Compared with Sycamore, Willow has twice the qubits, about five times the coherence, and much lower error.
• Compared with Quantinuum H2, Willow's gates are about a thousand times faster, but H2's qubits have higher fidelity and full connectivity.

Technical notes

Willow's key number is the error suppression factor, often written as lambda. When Google increased the surface code distance from 3 to 5 to 7, the logical error per cycle fell by a factor of roughly two at each step, giving a lambda of about 2. Because the factor is greater than one, adding more qubits makes the logical qubit better instead of worse, which is the defining property of operating below threshold.

Why it matters

Willow's result changed how people talk about scaling. Before it, adding qubits to an error-correcting code in a real machine might have made things worse. After it, the argument that larger codes give better logical qubits has direct experimental support on a superconducting platform.

It is also a reminder of how much work remains. Willow demonstrates a quantum memory that is better than its best physical qubit, but useful algorithms need logical gates, thousands of logical qubits, and error rates many orders of magnitude lower. Treat it as a milestone on the road, not the destination.

Frequently asked questions

Keep exploring

• Company profile: Google (Google Quantum AI)
• Related processor: Google Sycamore
• Related processor: IBM Heron (r1, r2, and r3)
• Related processor: Quantinuum H2

This profile is written for educational purposes and summarizes company announcements, published papers, and vendor specification sheets. Last reviewed: September 30, 2026. Figures marked as company-reported have not all been independently verified, and benchmarks differ between vendors, so compare them with care.

← Back to Hardware Database