NVIDIA released Ising Calibration 1.5, an open-source vision language model for automated quantum processor calibration. The model interprets diagnostic outputs and tunes quantum processing units without prior training examples, advancing AI-based QPU calibration.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
NVIDIA releases Ising Calibration 1.5, a 31B-parameter vision-language model that surpasses all open-source models in both zero-shot and in-context learning scenarios, and debuts an NVFP4 quantized version deployable on a single GPU or DGX Spark.
NVIDIA Ising Calibration 1.5 is a 31B-parameter vision-language model specifically designed to interpret diagnostic outputs from quantum processors and determine how to tune them for continued operation. Its core capability lies in analyzing unfamiliar diagnostic results without prior training examples, while also leveraging examples from related experiments for in-context learning.
This model is the latest version of the NVIDIA Ising model family. Previous generations already possessed basic calibration capabilities, but version 1.5 brings significant improvements in both zero-shot reasoning and in-context learning.
The model's training data comes from partner contributions and covers multiple qubit modalities, including superconducting qubits, quantum dots, ions, neutral atoms, and electrons on Helium, with a focus on calibration and control. This multi-modal training strategy enables the model to adapt to diagnostic requirements across different quantum hardware platforms.
Performance evaluation uses the QCalEval benchmark, which measures the model's ability to interpret experimental results, classify outcomes, assess significance, fit quality and key features, and recommend next steps. The evaluation covers both zero-shot and in-context learning scenarios.
On QCalEval, Ising Calibration 1.5 achieves an average of 10% higher zero-shot performance than the best open-source model of comparable size; in in-context learning, it improves 86.68% over the previous generation. The model performs best among open-source models and remains competitive with leading closed-source models such as Fable 5 and GPT 5.6 Sol.
Ising Calibration 1.5 is the first to offer an NVFP4 quantized version, which can run on a single GPU or NVIDIA DGX Spark, with model size reduced by 11.4% at BF16 precision. Throughput in tokens per second on DGX Spark is also optimized, supporting batching for parallelized calibration workflows across multiple experiments.
The model is suitable for data center GPUs such as NVIDIA Grace Blackwell and NVIDIA Vera Rubin, while the quantized version lowers the barrier for local deployment.
The NVIDIA Ising model family is fully open-source, with weights, data, benchmarks, and recipes provided, allowing users to modify, deploy, and fine-tune. Model weights are released on Hugging Face in BF16 and NVFP4 formats, and are also available as NVIDIA NIM and through NVIDIA Build.
For deployment, NVIDIA provides a quantum calibration agent blueprint based on the NVIDIA Nemo Agent Toolkit, enabling rapid setup of automated calibration workflows. The open-source license uses the Linux Foundation's OpenMDW, granting QPU builders and operators data control and flexible deployment capabilities.
This article's information primarily comes from NVIDIA's official technical blog, a first-party source. Performance data (e.g., 10% zero-shot improvement, 86.68% in-context learning improvement) are self-reported by NVIDIA and have not been independently verified by third parties. Factual information such as model parameters, quantized versions, and open-source licenses can be considered confirmed.
Ising Calibration 1.5, through multi-modal training and quantized deployment, pushes AI-driven quantum calibration from the lab to practical applications, and its open-source strategy may accelerate automated operations of quantum computing hardware.
Primary report
Primary source