NVIDIA open-sources Ising to speed fault-tolerant quantum work

NVIDIA is betting that one of quantum computing's ugliest bottlenecks looks increasingly like an AI problem. In its Apr 14 post, the company rolled out NVIDIA Ising as an open family of models for quantum processor calibration and quantum error correction decoding, two jobs that currently soak up human time and classical compute. That matters because the best quantum processors still fail roughly once in every thousand operations, while useful fault-tolerant systems need error rates closer to one in a trillion. If AI can compress part of that gap, the next limiting factor shifts from manual tuning toward deployable software.

The release is not a single model but two domains. Ising Calibration is a vision-language model for reading quantum experiment outputs and recommending next steps. Ising Decoding is a pair of 3D CNN pre-decoders for real-time error correction. NVIDIA says the package ships with open base models, training frameworks, deployment recipes, and fine-tuning and quantization workflows, letting labs adapt the stack to their own QPU noise characteristics while keeping proprietary data on-site. That open-stack angle is the bigger strategic move: it is easier for hardware teams to test AI assistance when weights, data recipes, and tooling are not trapped behind an API.

The strongest numbers in the post are on calibration. NVIDIA says its 35B Ising-Calibration-1 model was trained on partner data spanning superconducting qubits, quantum dots, ions, neutral atoms, and electrons on Helium. On the new QCalEval benchmark, the company reports average gains of 3.27% over Gemini 3.1 Pro, 9.68% over Claude Opus 4.6, and 14.5% over GPT 5.4. On the decoding side, the fast model has about 912,000 parameters and the accurate model about 1.79 million. NVIDIA says the fast pre-decoder plus PyMatching runs 2.5x faster and is 1.11x more accurate at d=13 and p=0.003, while the accurate version is 2.25x faster and 1.53x more accurate on the same setting.

There is still a long way between a benchmark post and a production quantum control stack, but NVIDIA is at least publishing the pieces that let others check the claim. The company says weights are on Hugging Face, training code is on GitHub under Apache 2.0 for decoder workflows, and the stack includes real-time APIs built on CUDA-Q QEC and CUDAQ-Realtime. One projected number stands out: with 13 GB300 GPUs and FP8 precision, NVIDIA says the fast model could reach 0.11 microseconds per round for 1000 rounds at surface code d=13. If outside labs can reproduce even part of that, Ising will matter less as a demo and more as a new software layer for fault-tolerant quantum work.