NVIDIA releases Ising Calibration 1.5 model

NVIDIA has introduced Ising Calibration 1.5, a 31-billion-parameter vision language model designed to automate the diagnosis and tuning of quantum processing units. This updated model achieves an 11.4 percent reduction in size at BF16 precision compared to its predecessor. For local lab deployments, NVIDIA has also released an NVFP4-quantized version. This quantization allows the model to run on a single graphics card or an NVIDIA DGX Spark system, matching the capabilities of leading closed models like Fable 5 and GPT 5.6 Sol.

The model is trained on diverse datasets spanning multiple qubit modalities, including superconducting qubits, quantum dots, ions, neutral atoms, and electrons on Helium. To evaluate its performance, researchers used the QCalEval benchmark, which tests a model's ability to interpret experimental results, classify outcomes, evaluate significance, assess fit quality, and recommend next steps. In zero-shot reasoning, Ising Calibration 1.5 scored an average of 10 percent better than the next best open model of comparable size. Furthermore, its in-context learning performance saw a massive 86.68 percent improvement over the previous version when analyzing diagnostic results alongside related experimental examples.

For quantum computing practitioners, this release simplifies the deployment of agentic calibration workflows directly within local laboratory environments. By utilizing the NVIDIA Nemo Agent Toolkit and the Quantum-Calibration-Agent-Blueprint, operators can quickly automate quantum processing unit bring-up and retuning tasks. The model's optimized tokens-per-second throughput on the DGX Spark enables parallelized workflows across multiple experiments per qubit at a fraction of the cost of cloud APIs. Distributed under the OpenMDW License, the full-parameter checkpoints, quantized versions, and open datasets are fully accessible, allowing operators to maintain complete data control while running on hardware ranging from consumer gaming cards to enterprise systems like NVIDIA Grace Blackwell and NVIDIA Vera Rubin.