NVIDIA open‑sources Ising

Quantum computers fail often enough that the best chips still make about one mistake every 1,000 operations, and NVIDIA on April 14 released open models meant to cut that failure problem in calibration and error correction. (nvidia.com) NVIDIA said the Ising family includes a 35 billion-parameter vision-language model for calibration, two smaller decoding models for quantum error correction, and public code on GitHub with model weights on Hugging Face. (github.com) (huggingface.co) Calibration is the tune-up step for a quantum processor: engineers read noisy measurement plots, adjust settings, and repeat until the qubits behave as expected. NVIDIA said its calibration model can read those plots and turn a process that takes days into one that takes hours. (developer.nvidia.com) Error correction is the second bottleneck: a classical computer has to spot and fix likely quantum errors before they pile up and ruin a calculation. NVIDIA said useful systems need error rates closer to one failure in a trillion operations, not one in a thousand. (developer.nvidia.com) The largest model, Ising-Calibration-1-35B-A3B, is built on Qwen3.5-35B-A3B and uses a mixture-of-experts design with about 35 billion total parameters and about 3 billion active per token. Its model card says it takes image-and-text input, returns technical text, and was released on April 14, 2026. (huggingface.co) NVIDIA introduced a new benchmark called QCalEval with 243 entries across 87 scenario types from 22 experiment families, covering superconducting qubits and neutral-atom systems. On that benchmark, the model card reports an overall score of 74.7 for Ising Calibration versus 55.5 for the Qwen3.5-35B base model, with scores averaged across GPT-5.4 and Gemini 3.1 Pro judges. (huggingface.co) NVIDIA’s press release said the decoding side of Ising is up to 2.5 times faster and 3 times more accurate than traditional approaches, and the company framed the software as a way to make hybrid quantum-classical systems more reliable. (nvidianews.nvidia.com) The company also 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 open release does not settle whether these models will generalize across quantum hardware outside NVIDIA’s tests, and the Hugging Face card says outputs should be validated by domain experts before acting on experimental conclusions. What NVIDIA has done, as of April 14, is put the models, datasets, and training workflow in public view so outside labs can test that claim on their own machines. (huggingface.co) (github.com)