NVIDIA has published Ising Calibration 1.5, an open-source vision-language model (VLM) designed to interpret diagnostic outputs from quantum processors (QPU) and recommend how to tune them to continue operating. The model targets the interpretation of plots, graphs and text from experiments to support QPU bring-up and retune workflows.
Capabilities of Ising Calibration 1.5
Ising Calibration 1.5 advances AI-based QPU calibration by handling unfamiliar diagnostic results in a zero-shot setting, while also accepting related experimental examples for in-context learning (ICL). At BF16 precision the model is 11.4% smaller, which helps deployment in local lab environments and reduces resource requirements for agentic calibration workflows.
Quantized variant and compute targets
For the first time NVIDIA offers an NVFP4-quantized build of Ising Calibration 1.5, enabling deployment on a single GPU or on NVIDIA DGX Spark, making it comparable with leading closed models such as Fable 5 and GPT 5.6 Sol. The full model contains 31 billion parameters and is intended for data-center GPUs such as NVIDIA Grace Blackwell and NVIDIA Vera Rubin. The NVFP4 quantized version allows running on a consumer gaming card or DGX Spark with only a small cost in accuracy. NVIDIA also optimized tokens-per-second (TPS) throughput on DGX Spark to provide good local performance at lower cost and better batching for parallelized agent workflows.
Evaluation on QCalEval
Ising Calibration 1.5 was trained on partner-contributed data spanning multiple qubit modalities, including superconducting qubits, quantum dots, ions, neutral atoms, electrons on Helium, and other systems specialized in calibration and control. Its performance is measured with the QCalEval benchmark, which evaluates a model's ability to interpret experimental results, classify outcomes, assess significance, evaluate fit quality and key features, and recommend next steps.
QCalEval tests both zero-shot reasoning (analyzing results independently) and in-context learning (evaluating results in the context of related samples). On this benchmark Ising Calibration 1.5 shows strong zero-shot performance and substantial gains when using related examples: the evaluation reports an approximate 86.68% improvement over the predecessor in ICL (one figure cited is 86.5%). In zero-shot tasks the model scores about 10% higher on average than the next-best open model of comparable size. Overall, Ising Calibration 1.5 outperforms all open models out of the box on QCalEval and remains competitive with state-of-the-art closed or 1T+ parameter models.
Availability, licenses and resources
The NVIDIA Ising model family is fully open: model weights, data, benchmarks and recipes are provided so users can modify, deploy and fine-tune models for their specific QPUs. Full-parameter checkpoints for Ising Calibration 1.5 are available on Hugging Face as:
- Ising Calibration 1.5 31B BF16
- Ising Calibration 1.5 31B NVFP4
Ising Calibration 1.5 is also distributed as an NVIDIA NIM and hosted through NVIDIA Build. The OpenMDW License from the Linux Foundation is offered to provide QPU builders and operators flexibility to retain data control and deploy anywhere.
Deployment guides and blueprints
NVIDIA provides a Quantum-Calibration-Agent-Blueprint GitHub repository that contains a ready-to-use agent harness blueprint for deploying Ising Calibration 1.5 with the NVIDIA Nemo Agent Toolkit. The blueprint supports building agentic workflows that automate quantum calibration experiments and can integrate large-model cloud APIs.
The QCalEval dataset and benchmark scripts are public so researchers and developers can run the benchmark and reproduce evaluations.
Summary
Ising Calibration 1.5 is a 31-billion-parameter open VLM tailored for interpreting QPU diagnostics and recommending calibration steps. With a smaller BF16 footprint, an NVFP4-quantized option for single-GPU or DGX Spark deployment, and strong QCalEval results in both zero-shot and in-context settings, the model aims to make AI-driven quantum calibration more accessible for lab-scale and data-center environments.
