NVIDIA announced Ising on April 14, timed to World Quantum Day, calling it the first family of open-source AI models built specifically for quantum processor calibration and error correction. The package includes a 35-billion-parameter vision-language model for automated calibration and a pair of compact 3D convolutional neural networks for real-time error decoding. Weights, training frameworks, and benchmark data are available on Hugging Face and GitHub.
Jensen Huang framed it in characteristically grand terms, but strip away the language about quantum-GPU supercomputers and you're left with something genuinely interesting: NVIDIA is trying to turn AI into the operating system layer for quantum hardware. And if it works, the lock-in implications are considerable.
A chatbot for your qubits
Ising Calibration is the stranger of the two models, and the more ambitious. It is a fine-tuned version of Qwen 3.5-35B-A3B (a mixture-of-experts VLM with about 3 billion active parameters per token) retrained on experimental data from quantum processors spanning superconducting qubits, quantum dots, neutral atoms, ions, and electrons on helium. The idea is that you feed it images of calibration experiment outputs, the kind of plots a physicist would normally stare at for hours, and it tells you what's wrong and what to adjust next.
Paired with NVIDIA's NeMo Agent Toolkit, Ising Calibration can run in an agentic loop: stream measurement data, interpret results, recommend adjustments, repeat. NVIDIA claims this reduces calibration time from days to hours. I'd want to see that claim tested across different hardware platforms before taking it at face value, but the concept is sound. Calibration is tedious, repetitive, and currently bottlenecked by the availability of experienced physicists. Automating even part of it would be a genuine contribution.
To evaluate the model, NVIDIA and its partners built QCalEval, which they describe as the first benchmark for agentic quantum calibration using real QPU data. The benchmark has 243 entries across 87 scenario types from 22 experiment families. On this test, Ising Calibration 1 outperformed Gemini 3.1 Pro by 3.27% on average, Claude Opus 4.6 by 9.68%, and GPT 5.4 by 14.5%, according to NVIDIA's technical blog.
Those margins against general-purpose models aren't shocking. You'd expect a domain-specific fine-tune to beat a generalist on domain-specific tasks. The more interesting question is whether the benchmark itself is representative, and whether those percentage improvements translate to faster calibration in practice. QCalEval tests six question types (technical description, conclusions, significance, fit quality, parameter extraction, success classification), which covers a lot of ground, but the benchmark is new and hasn't been independently validated.
The decoding models are tiny, and that's the point
Ising Decoding is architecturally simpler but solves a harder operational problem. Quantum error correction requires decoding syndrome measurements in real time, faster than errors accumulate. If your decoder can't keep up, your error correction budget falls apart. The current open-source standard is PyMatching, a minimum-weight perfect matching decoder.
NVIDIA's approach uses 3D CNNs as pre-decoders that handle localized syndrome errors before passing the remaining work to a global decoder like PyMatching. The fast variant has roughly 912,000 parameters; the accurate one, about 1.79 million. These are not large models. They're designed to run on GPUs alongside the quantum processor in real-time control loops, where every microsecond counts.
NVIDIA says the fast model plus PyMatching runs 2.5x faster than PyMatching alone and is 1.11x more accurate at code distance 13 and physical error rate 0.003. The accurate model is 2.25x faster and 1.53x more accurate under the same conditions. At higher code distances (d=31), NVIDIA claims the accurate model can deliver a 3x improvement in logical error rate, though that's based on training with d=13 data and extrapolating.
Those are NVIDIA's own numbers, tested on their own hardware (DGX GB300 for the pre-decoder, Grace Neoverse-V2 CPU for PyMatching). Independent reproduction would help.
"Open source" with an asterisk
The framing is "open source," but the licensing tells a more complicated story. The GitHub resources (training frameworks, cookbooks, deployment scripts) are released under Apache 2.0. The model weights, however, use the NVIDIA Open Model License. That's not the same thing, and teams adopting Ising need to read the fine print rather than assuming everything shares identical terms.
This matters because the whole release is structured to pull quantum computing teams deeper into NVIDIA's stack. Ising integrates with CUDA-Q (NVIDIA's hybrid quantum-classical computing platform), NVQLink (their QPU-GPU interconnect), cuQuantum, and NIM microservices. The models can run locally, which is good for data privacy, but "locally" still means on NVIDIA hardware.
The pattern should look familiar. Get the tools out early, make them free (or free-ish), let an ecosystem grow around your platform, and watch as switching costs compound. It's the CUDA playbook applied to quantum. And given that quantum hardware teams are small, resource-constrained, and rarely have ML expertise in-house, the appeal of a ready-made toolkit is obvious.
Who's already using it?
The press release lists an impressive roster of early adopters. Ising Calibration is being used by IonQ, Infleqtion, IQM, Atom Computing, Harvard's School of Engineering, Fermilab, Lawrence Berkeley National Lab, and the UK National Physical Laboratory, among others. Ising Decoding has been picked up by Cornell, Sandia National Laboratories, UC San Diego, the University of Chicago, and SEEQC.
That's a broad coalition of national labs, universities, and commercial quantum companies. Whether "using" means deep integration or early experimentation is unclear from the announcement, but the breadth of the list suggests NVIDIA did extensive partner outreach before the launch.
Quantum stocks reacted, with IonQ, Rigetti, and D-Wave all seeing double-digit rallies on the news. Market enthusiasm for quantum announcements tends to run ahead of the technology, but NVIDIA's involvement does carry a different weight than the typical quantum press release. When the company that built the infrastructure for the AI boom says it is building infrastructure for quantum, capital pays attention.
What's actually new here
Quantum error correction and calibration are real bottlenecks, and applying ML to them isn't a new idea. Research groups have been exploring neural network decoders for years. What NVIDIA is adding is packaging, scale, and ecosystem integration. The Ising models come with training frameworks, synthetic data generation tools, fine-tuning recipes, and deployment pipelines. For a quantum hardware team that knows qubits but not PyTorch, that's genuinely useful.
Whether this accelerates the timeline for practical quantum computing is a separate question. Analyst firm Resonance projects the quantum computing market will reach $11 billion by 2030. But market projections for quantum have been perennially optimistic. The best quantum processors still make errors roughly once every thousand operations; useful applications probably need error rates closer to one in a trillion. Ising is a tool for narrowing that gap, not closing it.
The models and frameworks are available now on Hugging Face, GitHub, and build.nvidia.com.
