Nvidia launches Ising models

Quantum computers store information in qubits, which are fragile enough that today’s best machines make about one error in every 1,000 operations. Nvidia said on April 14 it is releasing an open model family called Ising to automate two of the cleanup jobs that keep those systems running. (developer.nvidia.com) One job is calibration, which means tuning a quantum chip after reading noisy measurement plots, the way an engineer keeps retuning a sensitive instrument. Nvidia said its Ising Calibration model is a 35 billion-parameter vision-language model trained to read those plots and suggest the next tuning step. (developer.nvidia.com; research.nvidia.com) The other job is decoding, which means spotting and correcting errors fast enough that they do not pile up before the qubits lose their state. Nvidia said its Ising Decoding release includes task-specific models and training tools, with one decoder running up to 2.5 times faster and another reaching 3 times the accuracy of traditional approaches in the company’s tests. (nvidianews.nvidia.com; developer.nvidia.com) The numbers matter because fault-tolerant quantum computing needs error rates far below what current hardware delivers. Nvidia said useful systems will need error rates closer to one in a trillion operations, not one in a thousand, which turns calibration and decoding into constant, high-speed control problems. (developer.nvidia.com) Nvidia has been building toward this for more than a year through its CUDA-Q quantum software stack and its DGX Quantum reference architecture with Quantum Machines. In March 2025, the company said DGX Quantum could connect graphics processing units to quantum hardware with round-trip latencies below 4 microseconds for calibration, control and decoding. (developer.nvidia.com; developer.nvidia.com) Nvidia is calling Ising “open,” but the release is not a general-purpose chatbot-style model and it does not stand alone from Nvidia’s own stack. The company’s Ising page says the models ship with pretrained weights, data, retraining guidance and deployment tooling tied to CUDA-Q and Nvidia’s quantum platform. (nvidia.com; developer.nvidia.com) The calibration side also comes with a new benchmark called QCalEval, which Nvidia researchers published on April 14 with 243 samples across 87 scenario types from 22 experiment classes. Nvidia said the benchmark is meant to test how well vision-language models interpret the plots quantum engineers already use to tune chips. (research.nvidia.com) Nvidia’s pitch is that quantum hardware companies can use graphics processing units not just to simulate quantum systems, but to run the real-time software wrapped around them. The first Ising releases keep that focus on the two repetitive jobs that sit between a noisy lab device and a fault-tolerant machine. (developer.nvidia.com; developer.nvidia.com)