What IBM announced
IBM has introduced IBM Quantum Nighthawk r3, the next superconducting quantum processor in the Nighthawk family. According to the source, the chip is designed to perform 10,000 quantum operations with accuracy on 520 programmable qubits, distributed across 2,034 physical qubits and connected by 10 interconnects called l-coupler. The company plans to combine two r3 chips into a single 1,040-qubit system in 2027.
The key point to keep clear is the distinction between product and plan: the r3 is an announced design, not hardware already shipping. The 1,040-qubit system belongs to the future roadmap and is not an available result.
The installed base: how Nighthawk r2 works
The chip currently in production is Nighthawk r2, with 120 programmable qubits, 218 couplers and 120 reset elements, for a total of 458 physical quantum elements. The Nighthawk family uses a square lattice in which most qubits connect to four neighbors, versus two or three in previous architectures; this connectivity allows circuits to express more entanglement structure with fewer swap operations.
The most significant technical change concerns qubit reset. The previous conditional approach does not correct qubits that have leaked out of the expected state, forcing the chip to wait hundreds of microseconds between executions. Nighthawk r2 introduces a dissipative mechanism per qubit that, via a tunable coupler, connects the qubit to a cold environment on demand: the effective T1 time drops from about 200 microseconds to about 25 nanoseconds, and the wait between executions can be reduced to as little as one microsecond. This yields the claimed 100,000 circuits per second, versus about 4,000 for the Heron chip, and an initialization error reported as 25 times lower.
Claimed performance and open questions
IBM reports that customers have run 34 billion circuits on Nighthawk r2. The cited speed gains are self-reported by the company and its partners: E.ON reports a 15-fold acceleration in quantum machine learning for optimizing energy networks compared with Heron; Oak Ridge National Laboratory indicates a 12-fold gain on a simulation of magnetic materials; Algorithmiq reports circuits executed 8 times faster; Q-CTRL describes a 60-qubit fermionic simulation with 4,497 two-qubit gates, reduced from 4 minutes and 25 seconds to 23 seconds.
Elements remain to be verified. Inter-chip links have achieved an average state-transfer fidelity of 99%, with a maximum of 99.3%, below the 99.9% target. Moreover, the r3, including its use in 1,040-qubit systems, should be considered a project: its actual availability and real-world performance have not yet been demonstrated. The source also cites the completion of a 2026 roadmap milestone concerning the accurate estimation of observables on circuits with more than 7,500 gates, but this is an IBM statement.
