NVIDIA releases Ising for qubit calibration
NVIDIA announced Ising, an open AI model aimed at qubit calibration for quantum computing that includes a vision‑language component. The release is presented as a bridge between machine learning and emerging quantum hardware tooling. (x.com)
NVIDIA said on April 14 that it released Ising, an open family of artificial intelligence models for quantum computing, starting with qubit calibration and error correction. (nvidianews.nvidia.com) A qubit is the basic unit of a quantum processor, and it is fragile enough that today’s best systems make about one error in every thousand operations, according to NVIDIA’s technical blog. Ising is meant to help tune those devices and catch errors fast enough for the hardware to stay usable. (developer.nvidia.com) The first calibration model, Ising Calibration 1, is a 35 billion parameter vision-language model trained on multimodal qubit data. NVIDIA said it reads the plots and measurements quantum labs already use and can plug into an agentic workflow that automates processor bring-up and re-tuning. (developer.nvidia.com) Calibration is the routine of measuring a quantum chip, adjusting control signals such as microwaves or lasers, and repeating that loop until the device behaves as expected. NVIDIA said Ising can cut that process from days to hours by interpreting quantum processor measurements and suggesting the next action. (nvidianews.nvidia.com) NVIDIA also released Ising Decoding, a separate set of models for quantum error correction, the classical sidecar system that spots and fixes quantum mistakes before they pile up. The company said those decoder models run up to 2.5 times faster and reach up to 3 times higher accuracy than traditional approaches. (nvidianews.nvidia.com) The company paired the model release with a new benchmark called QCalEval, which tests whether vision-language models can understand quantum calibration plots. NVIDIA researchers said the benchmark contains 243 samples across 87 scenario types from 22 experiment families, covering superconducting qubits and neutral-atom systems. (research.nvidia.com) In that benchmark, NVIDIA said Ising Calibration 1 reached a 74.7 zero-shot average score. The developer site says that put it 3.27 percent ahead of Gemini 3.1 Pro, 9.68 percent ahead of Claude Opus 4.6, and 14.5 percent ahead of GPT 5.4 on the company’s calibration test suite. (research.nvidia.com) (developer.nvidia.com) NVIDIA is distributing Ising through a public GitHub repository, Hugging Face model pages, and a deployable NVIDIA Inference Microservices endpoint. The GitHub landing page lists an Apache 2.0 license for the repository and links to the calibration model, two decoder models, datasets, and cookbooks. (github.com) (developer.nvidia.com) NVIDIA said early adopters include Academia Sinica, Fermi National Accelerator Laboratory, Harvard John A. Paulson School of Engineering and Applied Sciences, Infleqtion, IQM Quantum Computers, Lawrence Berkeley National Laboratory’s Advanced Quantum Testbed, and the United Kingdom’s National Physical Laboratory. (nvidianews.nvidia.com) The release gives quantum hardware teams a new pitch for using large models less as chatbots and more as lab software that reads instrument output, recommends settings, and feeds classical control loops. NVIDIA’s framing is that useful quantum machines will need that kind of classical automation before fragile qubits can scale into reliable systems. (developer.nvidia.com)
