AI boosts quantum computing! NVIDIA open-sourced the 31B visual model Ising 1.5, fully automatic calibration for quantum computers

AI boosts quantum computing technology for a major breakthrough! NVIDIA’s technical blog, released today (27th), says that NVIDIA has officially open-sourced the 31 billion-parameter visual language model “Ising Calibration 1.5,” designed specifically for diagnosing and tuning quantum processing units (QPUs). The model not only supports multiple qubit technologies, but also adds an NVFP4 quantized version, allowing labs to deploy fully automated quantum computer calibration agents easily in local environments with much lower hardware requirements.
(Background: Smashed $500 billion! NVIDIA teams up with OpenAI to build a “world’s largest” AI data center, Japan throws big money into the mix)
(Background addition: Soccer king Messi also backs world models! His foundation invests in Fei-Fei Li’s startup World Labs, joining forces with Nvidia and AMD)

As cutting-edge AI technology makes rapid strides, AI is helping humans solve the most complex physical computation problems. On July 27, according to Taipei time, NVIDIA’s research team officially released “NVIDIA Ising Calibration 1.5.” This is a visual language model (VLM) with 31 billion parameters (31B). Its only and focused task is to interpret the diagnostic outputs of quantum processing units (QPUs) and decide how to tune them to keep the system operating stably.

ICL performance surges by 86%; quantized version lowers deployment barriers

This new 1.5 model achieves dual breakthroughs in both performance and lightweight deployment. Based on results from the latest QCalEval benchmark tests, Ising Calibration 1.5 demonstrates state-of-the-art zero-shot and in-context learning (ICL) capabilities. When analyzing previously unseen diagnostic results, its zero-shot performance is on average about 10% higher than other open-source models of the same size. Meanwhile, in the ICL tests using relevant samples, performance improves dramatically by 86.68% compared with the previous-generation model. Overall strength is already able to compete with top closed-source models at the 1T-parameter scale (such as Claude Fable 5, etc.).

To make adoption easier for research organizations, NVIDIA is providing, for the first time, an NVFP4 quantized version, and it also successfully reduced the model size by 11.4% under BF16 precision. This means that QPU diagnostic models, which previously required massive compute power, can now run smoothly with a single GPU or NVIDIA DGX Spark—greatly lowering the hardware barrier for labs to deploy agentic calibration workflows locally.

Supports multiple quantum modalities; fully open-sourced resources for sharing

In terms of training data, Ising Calibration 1.5 draws on real QPU data from many partner organizations, covering multiple leading qubit technologies, including: superconducting qubits, quantum dots, ion traps, neutral atoms, and electrons on helium. This gives the model high generality, enabling precise evaluation of various fitting qualities and key features, and recommending the next calibration actions.

Aligned with the goal of advancing quantum computing worldwide, NVIDIA has fully open-sourced the complete parameter checkpoints, quantized versions, and deployment blueprints under the Linux Foundation’s OpenMDW License. Currently, developers can directly obtain model weights on the Hugging Face platform, or quickly build via NVIDIA NIM containers and hosted services. This technology’s deployment is expected to give a strong push to the automation and scaling of quantum computers worldwide.