NVIDIA opens Ising models

Quantum computers are fragile machines, and NVIDIA has released an open-source model family called Ising to help keep them tuned and corrected in real time. (nvidianews.nvidia.com) NVIDIA announced Ising on April 14, 2026, as a set of pre-trained models, training frameworks, datasets, benchmarks, and deployment recipes for quantum processor calibration and quantum error correction. The company said the models are available through its developer site, GitHub, Hugging Face, and NVIDIA Inference Microservices. (nvidianews.nvidia.com) (developer.nvidia.com) (github.com) Calibration is the process of tuning a quantum chip so its qubits behave as intended, like adjusting thousands of tiny dials that drift over time. Error correction is the process of spotting and fixing mistakes fast enough that a quantum computation does not fall apart before it finishes. (nvidia.com 1) (nvidia.com 2) NVIDIA’s first release includes Ising Calibration, a vision-language model for automating quantum device tuning, and Ising Decoding, two small neural-network decoders for surface-code error correction, one tuned for speed and one for accuracy. NVIDIA’s GitHub page lists the calibration model at 35 billion parameters with 3 billion active parameters, and the two decoder variants at 1.79 million and 0.91 million parameters. (nvidia.com) (github.com) The company said Ising Decoding delivers up to 2.5 times the speed and 3 times the accuracy of pyMatching, a widely used open-source decoder benchmark. NVIDIA also said the models can be retrained, fine-tuned, quantized, and deployed on hardware close to the quantum processor. (developer.nvidia.com) (github.com) The release extends a line of NVIDIA quantum work that has focused on pairing quantum processors with classical graphics processing units, or GPUs, that handle simulation, control, and low-latency feedback. In March 2025, NVIDIA and Quantum Machines said their DGX Quantum reference architecture connected GPUs to quantum hardware with round-trip latencies below 4 microseconds. (developer.nvidia.com) NVIDIA had already been testing artificial intelligence decoders with hardware partners before this launch. In March 2025, the company said a transformer-based decoder built with QuEra could be trained largely on synthetic data from simulations rather than only on measurements from physical machines. (developer.nvidia.com) The open-source framing matters because quantum hardware companies often treat calibration data, control software, and decoder pipelines as proprietary infrastructure. NVIDIA said Ising is fully open, with weights, data, benchmarks, and recipes released so outside researchers can modify the models for their own quantum processing units. (developer.nvidia.com) NVIDIA’s pitch is that artificial intelligence will sit in the control loop of quantum machines, not just in chatbots or coding tools. Jensen Huang said in the launch announcement that “AI is essential to making quantum computing practical,” tying the company’s model strategy directly to specialized scientific computing. (nvidianews.nvidia.com) What happens next is less about consumer software than lab adoption: whether quantum hardware teams use Ising’s open models as a starting point for their own chips, and whether NVIDIA’s claimed gains hold up outside its benchmarks. For now, the company has turned quantum calibration and error correction into another front in the race to make AI infrastructure indispensable. (nvidia.com) (developer.nvidia.com)