AI has calibrated all 20 qubits on an IQM quantum chip. It read every tuning chart, decided what to do next and explained each decision in plain English. It runs on open models from NVIDIA, and it is measured on a new open benchmark that IQM helped build.
It’s the next chapter of the agentic calibration work using NVIDIA Ising that we announced in April.
Why it matters: before a quantum computer can do useful work, every qubit has to be tuned. Each tuning step produces a chart, and today an engineer reads it and decides what to do next. As processors grow, that becomes the bottleneck. Quantum engineers are scarce, and the number of charts keeps climbing.
So we asked a simple question: can an AI read these charts as well as an engineer? Here’s what we found.
The models are vision language models. They take an image and text together, so they can look at a chart and explain what they see.
General-purpose models of this kind have seen millions of photos and documents, but very few calibration charts. That is the gap.
NVIDIA Ising is a family of open models for quantum computing. Leveraging NVIDIA Ising Calibration in an agentic workflow helps quantum experts automate quantum processor bring-up and retune calibration workflows without requiring machine learning domain expertise.
Two things make them practical for the labs and supercomputing centers that run quantum systems:
Until recently there was no common way to measure how well an AI reads calibration charts, so there was no way to compare one model with another.
That has changed. NVIDIA built QCalEval, the first open benchmark for the job, with IQM and other quantum collaborators.
IQM contributed chart data from our chips and co-wrote the paper. There’s a public leaderboard, so anyone can test a model and see where it lands.
The result: on QCalEval, the NVIDIA Ising Calibration models score up to 19 points higher than alternatives. For this job, specialized beats general.
Why this matters: a shared, open test is how a field tracks real progress. Without one, every claim about AI calibration is a demo on someone’s own data. With one, anyone can check a result, compare models and see whether the next version is actually better.
Then we put it to work. We connected the models to IQM’s own calibration software and ran them on a 20-qubit IQM chip.
The AI read each chart and decided the next step. In several steps, the right setting for some qubits was off the edge of the chart. The AI spotted which qubits, widened the range and reran only those. That is exactly what an experienced engineer would do.
All 20 qubits ended up calibrated. Every decision is written down in plain English, so an engineer can check what the AI did and why.
This was 20 qubits and problems an engineer would recognize. Three things come next:
The goal hasn’t changed since April: quantum computers that calibrate themselves and explain what they did, so they can run in AI factories and supercomputing centers without a resident team of quantum specialists.
Max Haeberlein leads software ecosystem partnerships at IQM Quantum Computers. He owns the technical relationships that connect IQM’s hardware to the wider HPC and AI software stack. Before IQM he spent seven years at Intel, building AI applications for chip design and leading the team behind them. He started at the Walther-Meissner-Institute, working on some of Germany’s early superconducting quantum processors. Max works where quantum hardware, software, and commercial strategy meet.
Search faster—hit Enter instead of clicking.
Click to expand
