NVIDIA Ising Calibration 1.5 Automates Quantum Tuning with Open VLM

NVIDIA released Ising Calibration 1.5, an open source vision language model that reads diagnostic outputs from quantum processors and recommends tuning actions. The 31-billion-parameter model is trained on partner data across multiple qubit modalities, including superconducting qubits, quantum dots, ions, neutral atoms, and electrons on Helium. It is 11.4% smaller at BF16 precision than its predecessor, easing deployment of agentic calibration workflows in local lab environments.

Performance and evaluation

NVIDIA evaluates the model using the QCalEval benchmark, which measures the ability to interpret experimental results, classify outcomes, evaluate significance, assess fit quality and key features, and recommend next steps. The evaluation covers both zero-shot reasoning, which analyzes results independently, and in-context learning (ICL), which uses examples from related experiments when available.

On QCalEval, Ising Calibration 1.5 shows strong zero-shot performance without prior examples. With ICL, it improves 86.68% over its predecessor. NVIDIA states it outperforms comparable open models and remains competitive with leading closed models, including those with 1 trillion or more parameters. For zero-shot, the model scores 10% better on average than the next best open model at comparable size.

Deployment options

The full BF16 model targets data center GPUs such as NVIDIA Grace Blackwell and NVIDIA Vera Rubin. For the first time, a version quantized to NVFP4 is available, enabling deployment on a single GPU or an NVIDIA DGX Spark. NVIDIA says the quantized version has only a small cost in accuracy. Tokens per second throughput on DGX Spark has been optimized, and the model supports batching, which allows parallelizing the agent across multiple experiments for each qubit.

Model weights, datasets, benchmarks, and deployment recipes are released under the OpenMDW License from the Linux Foundation. Hugging Face checkpoints are available for both BF16 and NVFP4 precision. A Quantum Calibration Agent Blueprint, available in the NVIDIA/Quantum-Calibration-Agent-Blueprint GitHub repository, provides a script for deploying an agentic workflow using the NVIDIA Nemo Agent Toolkit. The QCalEval benchmark script and dataset are also released. The model is available as an NVIDIA NIM and hosted through NVIDIA Build.