NVIDIA open‑sources Ising models

NVIDIA on April 14 released Ising, an open-source set of artificial intelligence models for tuning quantum chips and correcting their errors. (nvidia.com) The release includes two starting points: Ising Calibration, a vision-language model for quantum device tuning, and Ising Decoding, a pair of neural-network decoders for surface-code error correction. NVIDIA published the models, datasets, cookbooks, and deployment tooling through its website and GitHub repositories. (developer.nvidia.com) Quantum computers use qubits, which are fragile bits that can drift out of tune or flip by mistake when heat, noise, or control errors creep in. NVIDIA said useful systems need error rates to fall from about one in a thousand operations to roughly one in a trillion. (developer.nvidia.com) Calibration is the step where engineers tune a quantum processor’s control settings, like adjusting thousands of tiny knobs until the chip behaves as intended. NVIDIA’s main calibration model, Ising-Calibration-1-35B-A3B, is listed as a 35-billion-parameter vision-language model and is available on Hugging Face and as a deployable NVIDIA Inference Microservice. (github.com) Error correction is the step where software reads noisy signals from qubits and guesses which mistakes happened before the computation collapses. NVIDIA said its Ising Decoding models run up to 2.5 times faster and reach 3 times the accuracy of PyMatching, a widely used open-source decoder baseline. (nvidianews.nvidia.com) The public GitHub landing page shows three launch models: one calibration model and two decoder variants called Fast and Accurate. The decoder checkpoints are listed at about 0.91 million and 1.79 million parameters, and NVIDIA published a separate training framework for retraining and deployment. (github.com; github.com) NVIDIA tied the release to “agentic” workflows, meaning software agents can inspect calibration images, choose actions, and iterate without a human hand-tuning every step. The company also published a QCalEval dataset described as an evaluation set for quantum calibration agents. (github.com) The launch extends a quantum software push NVIDIA had already been building around CUDA-Q and DGX Quantum, which connect graphics processors to quantum hardware and simulators. In a 2025 post, NVIDIA said its quantum error-correction stack already included tools for generating training data and accelerating decoders on graphics processors. (developer.nvidia.com) NVIDIA framed Ising as infrastructure for researchers and hardware companies rather than a finished quantum computer product. The company’s developer page says the aim is to give quantum specialists access to modern machine-learning tools without requiring them to become machine-learning experts first. (developer.nvidia.com) The immediate test is whether outside labs adopt the open models, retrain them on their own devices, and match NVIDIA’s benchmark claims. For now, the company has put its pitch in public code: artificial intelligence as a support system for building quieter quantum machines. (github.com; nvidia.com)