Quantum computers have a persistent maintenance problem that doesn’t get talked about much in the coverage of qubit counts and gate fidelities: the hardware drifts constantly, and keeping it calibrated is expensive, slow, and expert-intensive.

Superconducting qubits, the dominant architecture from IBM, Google, and most startups, are physically realized as tiny oscillating circuits cooled to millikelvin temperatures. Their frequencies and coupling strengths shift continuously due to charge noise, flux noise, and material imperfections. A processor that was calibrated yesterday might need adjustment today. The calibration procedure involves running characterization experiments, analyzing the output, and tweaking control parameters — a loop that has historically required significant hands-on time from physicists who understand the specific hardware.

On April 14, NVIDIA released Ising: two open-source model families designed to automate these problems.

Ising Calibration

The calibration model is a 35-billion-parameter vision-language model fine-tuned to read experimental measurements from a quantum processing unit (QPU) and infer the parameter adjustments needed to tune it. The “vision” part matters: QPU measurements are often rendered as 2D color plots (chevron patterns, spectroscopy sweeps, randomized benchmarking curves), and a VLM can interpret these diagnostic images the way a physicist would — reading the shape of a resonance feature and concluding which control knobs need adjustment.

Paired with an agent that runs experiments and feeds results back to the model in a loop, calibration time drops from days to hours. The model is available on Hugging Face and build.nvidia.com.

The design choice to use a VLM is interesting. An alternative would be a specialized regression model trained purely on numerical measurement vectors. VLMs are heavier and slower, but they transfer better: a model trained to read calibration plots for superconducting qubits can generalize to trapped-ion or photonic processors because the visual grammar of experimental plots is more universal than the specific numerical formats each hardware type uses. It also means the model benefits from pre-training on the enormous corpus of physics and experimental science figures that are part of any large multimodal model’s training data.

Ising Decoding

The second family addresses quantum error correction, which is a different but equally critical infrastructure problem.

Modern quantum computing at scale requires error correction: qubits are physically noisy, so you encode logical qubits across many physical qubits and use a protocol (typically a surface code or similar stabilizer code) to detect and correct errors without measuring the logical state. The bottleneck is decoding: you run stabilizer measurements to detect error syndromes, then decode those measurements to figure out what correction to apply. This has to happen in real time, faster than the qubits decohere between rounds.

Ising Decoding consists of two 3D convolutional neural network variants — one optimized for speed, one for accuracy — that function as pre-decoders running alongside pyMatching, the current open-source MWPM standard. The pre-decoders handle a fast, approximate first pass over syndrome patterns, resolving the majority of errors before forwarding residual cases to pyMatching for final decoding.

The “3D” structure reflects that error syndrome data is naturally three-dimensional: two spatial dimensions from the qubit lattice, plus time across measurement rounds. A decoder needs to reason about correlated errors across this full space-time volume to correctly infer the error chain. The small model sizes — roughly 912,000 parameters for the Fast variant and 1.79M for Accurate — mean the pre-decoding step adds minimal latency.

The benchmark numbers on surface code decoding: the combined system runs up to 2.5× faster and up to 3× more accurate than pyMatching alone. The performance gap matters practically: in a real fault-tolerant system, the decoder must keep up with the physical measurement rate, and any throughput headroom translates directly to lower overhead and longer time margins before decoherence bites.

Why this is notable

The model names reference the Ising model — a statistical mechanics framework for interacting spins that’s historically connected to both quantum systems and computational problems. The naming isn’t just flavor: Ising model problems map naturally onto error correction (finding the minimum energy configuration corresponds to finding the most likely error chain).

The broader claim — that AI is becoming “the control plane for quantum machines” — is a real trend. The calibration loop has always been software-driven; what’s new is that the analysis step that required physicist expertise is now being automated via learned models. This matters more as quantum hardware scales: a 100-physical-qubit system can be calibrated by a small team; a 10,000-qubit system probably cannot without automation.

Adoption is already underway at institutions running real quantum hardware: Harvard, Fermilab, Lawrence Berkeley National Lab’s Advanced Quantum Testbed, NPL, IQM, IonQ, Infleqtion, and Atom Computing are among the named early users. These aren’t press-release-only partnerships — institutions running actual QPUs have limited patience for tools that don’t work.

The models are open source on GitHub and Hugging Face, integrated with NVIDIA’s CUDA-Q platform. Whether you have a superconducting, trapped-ion, or neutral-atom processor, the calibration and decoding challenges are similar enough that the same model families can address them across hardware platforms.