NVIDIA Ising Calibration 1.5 发布：310 亿参数 VLM 实现全自动量子计算机校准

Technical Blog

sessionStorage.removeItem('facets.search.text'); sessionStorage.removeItem('facets.search.text-copy'); sessionStorage.removeItem('facets.search.filters');

Jul 27, 2026

NVIDIA Ising Calibration 1.5 is an open-source vision language model trained on partner data across multiple qubit modalities to interpret quantum processor diagnostics and recommend tuning actions. The 31-billion-parameter model achieves strong zero-shot performance on the QCalEval benchmark and improves 86.68% over its predecessor when using in-context learning with related experimental examples. An NVFP4-quantized version enables deployment on a single GPU or NVIDIA DGX Spark with optimized tokens-per-second throughput, while the full BF16 model targets data center GPUs such as NVIDIA Grace Blackwell and NVIDIA Vera Rubin. Model weights, datasets, benchmarks, and deployment recipes are released under the OpenMDW License, including Hugging Face checkpoints for both BF16 and NVFP4 precision, a Quantum Calibration Agent Blueprint, and the QCalEval benchmark script. Next Steps Download the Ising Calibration 1.5 31B BF16 model weights from Hugging Face. Download the Ising Calibration 1.5 31B NVFP4 quantized model weights from Hugging Face. Explore the Quantum Calibration Agent Blueprint to deploy an agentic calibration workflow.

Powered by NVIDIA Nemotron. AI-generated content may summarize information incompletely. Verify important information. Learn more

NVIDIA Ising Calibration is an open source vision language model (VLM) designed to interpret diagnostic outputs from quantum processors and determine how they should be tuned to continue operating.

This post introduces the latest model release, NVIDIA Ising Calibration 1.5 , which advances AI-based QPU calibration by analyzing unfamiliar diagnostic results without prior training examples. Ising Calibration 1.5 also uses examples from related experiments when available and is 11.4% smaller at BF16 precision. This eases the deployment of agentic calibration workflows directly in local lab environments.

For the first time, the model is also available in an NVFP4 -quantized version, enabling deployment on a single GPU or an NVIDIA DGX Spark—comparable with leading closed models such as Fable 5 and GPT 5.6 Sol.

The Ising Calibration 1.5 model is trained on data generated from partner contributions across multiple qubit modalities, including superconducting qubits, quantum dots, ions, neutral atoms, electrons on Helium, and others specializing in calibration and control.

Performance of Ising Calibration 1.5 is evaluated using the QCalEval benchmark , which measures a model’s ability to interpret experimental results, classify outcomes, evaluate significance, assess fit quality and key features, and recommend next steps. For additional details on the benchmark, model architecture, and evaluation results, see QCalEval: Benchmarking Vision-Language Models for Quantum Calibration Plot Understanding .

The evaluation covers both zero-shot and in-context learning (ICL). Zero-shot reasoning analyzes results independently, while ICL evaluates results in the context of related samples. Both are important for building agents that can automate QPU bring-up and retune operations.

On the QCalEval benchmark, Ising Calibration 1.5 shows strong performance when analyzing diagnostic results without prior examples. It is now also 86.68% better than its predecessor when using examples from related experiments. It outperforms comparable open models and remains competitive with leading closed models.

Ising Calibration 1.5 advances AI and quantum computing calibration by outperforming all open models out of the box and on the QCalEval benchmark. It is competitive with state-of-the-art closed or 1T+ parameter models.

Figure 1. For zero-shot, Ising Calibration 1.5 scores 10% better on average than the next best open model at comparable size. For ICL, Ising Calibration 1.5 is 86.5% better than its predecessor

The 31-billion-parameter VLM is suited for data center GPUs such as NVIDIA Grace Blackwell and NVIDIA Vera Rubin. It also ships with a quantized version to NVFP4 where users can run this model on a consumer gaming card or NVIDIA DGX Spark with only a small cost in accuracy.

The tokens per second (TPS) performance on DGX Spark has also been optimized, enabling good throughput locally for a fraction of the cost.

Figure 2. Ising Calibration 1.5 provides better performance on DGX Spark. It stands out with batching, enabling fast workflows for users looking to parallelize their agent across multiple experiments for each qubit

To learn more about deploying Ising Calibration 1.5 with an agent, check out the NVIDIA/Quantum-Calibration-Agent-Blueprint GitHub repo.

The NVIDIA Ising model family is fully open. Weights, data, benchmarks, and recipes are provided so you can modify, deploy, and fine-tune your own models and variants for your specific QPUs.

Full-parameter checkpoints for Ising Calibration 1.5 are available on Hugging Face:

Ising Calibration 1.5 is also available as an NVIDIA NIM and hosted through NVIDIA Build . The OpenMDW License from Linux Foundation offers QPU builders and operators with the flexibility to maintain data control and deploy anywhere.

A ready-to-use agent harness blueprint is available with support for this model and others, including large model cloud APIs.

Ising Calibration 1.5 is built on real QPU data provided by partners and collaborators. A semantic quantum calibration benchmark has also been released to evaluate model effectiveness for this task.