Nvidia Releases Ising, an Open-Source AI Model Family Aimed at Making Quantum Computers Actually Work
Nvidia's new Ising family targets quantum error correction and calibration, claiming 2.5x faster and 3x more accurate decoding than the current open-source standard, with more than 20 labs and hardware makers already on board.
Overview
Nvidia has released Ising, a family of open-source AI models designed to tackle two of the hardest problems standing between today’s noisy quantum processors and genuinely useful quantum computers: error correction and calibration. The company announced the launch on April 14, with the models distributed through GitHub, Hugging Face, and build.nvidia.com.
The release lands as the quantum industry is converging on a consensus that classical AI, not just better qubits, will be needed to keep error rates in check long enough for quantum advantage to materialize. Nvidia’s framing of the launch is unusually direct: in a statement accompanying the announcement, chief executive Jensen Huang said that “AI is essential to making quantum computing practical” and that “with Ising, AI becomes the control plane the operating system of quantum machines.”
What Ising Actually Contains
The Ising family has two components. The first, Ising Calibration, is a 35-billion-parameter vision-language model that reads experimental measurements from quantum processors and infers the calibration adjustments needed to keep the hardware stable, as described by Tom’s Hardware. Nvidia says the model compresses calibration cycles that currently take engineering teams days down to hours, according to its press release.
The second component, Ising Decoding, addresses the real-time decoding bottleneck in quantum error correction. It ships as two variants of a 3D convolutional neural network with 0.9 million and 1.8 million parameters, optimized respectively for speed and accuracy, and performs pre-decoding for surface-code error correction, Tom’s Hardware reported. Nvidia benchmarks Ising Decoding against pyMatching, the incumbent open-source decoder, and claims it is up to 2.5 times faster, 3 times more accurate, and needs roughly 10 times less training data than competing tools, per Tom’s Hardware.
Why those numbers matter is a point Nvidia’s quantum product director Sam Stanwyck made explicit to Tom’s Hardware: current quantum processors produce an error roughly once every thousand operations, and “the logical error rate is directly tied to how quickly decoding runs alongside the hardware.” A decoder that cannot keep up with the stream of syndrome measurements from a running quantum chip forces the system to either slow down or accumulate unrecoverable errors.
Who Is Already Using It
The partner list that Nvidia published alongside the launch is notable for spanning most of the hardware modalities in the field. According to the Nvidia newsroom, Ising Calibration is already being deployed or evaluated by Atom Computing, Academia Sinica, Conductor Quantum, EeroQ, Fermi National Accelerator Laboratory, Harvard’s John A. Paulson School of Engineering and Applied Sciences, Infleqtion, IonQ, IQM Quantum Computers, Lawrence Berkeley National Laboratory’s Advanced Quantum Testbed, Q-CTRL, and the U.K. National Physical Laboratory.
Ising Decoding users listed in the same announcement include Cornell University, EdenCode, Infleqtion, IQM, Quantum Elements, Sandia National Laboratories, SEEQC, UC San Diego, UC Santa Barbara, the University of Chicago, the University of Southern California, and Yonsei University. That roster covers superconducting, trapped-ion, neutral-atom, and photonic architectures, which means the models are being tested against fundamentally different error profiles rather than a single platform.
The timing of the release is not accidental. The Quantum Insider notes that Nvidia staged the announcement around World Quantum Day, and cites an estimate from analyst firm Resonance that the quantum computing market will exceed $11 billion by 2030.
Open Source, With Caveats
Although the Ising weights are being distributed under an open-source license on GitHub, Hugging Face, and build.nvidia.com, the broader control-plane stack that Nvidia is positioning around them is not. Tom’s Hardware observes that the NVQLink interconnect, the CUDA-Q software platform, and the deployment tooling designed to run Ising alongside live quantum processors remain Nvidia-exclusive. In practice, a lab that wants the fastest path to the advertised performance will end up running the open models on top of a proprietary Nvidia stack.
That mirrors a pattern Nvidia has used before in classical AI: open weights at the top of the pyramid, proprietary infrastructure underneath. Whether it produces the same effect in quantum computing where several of the hardware vendors on the partner list already run their own classical control systems will depend on how tightly the error-correction loop needs to be coupled to Nvidia silicon to hit the published benchmarks.
What We Don’t Know
Nvidia has not published an independent third-party reproduction of the 2.5x speed and 3x accuracy claims for Ising Decoding, and the 10x training-data figure cited by Tom’s Hardware comes from Nvidia’s own benchmarks against pyMatching. It is also unclear how the models perform on hardware architectures that do not use surface codes, and the press materials do not specify which noise models were used during training. Partners on the published list have not yet released their own evaluations of the models in production use.
This launch extends The Machine Herald’s earlier coverage of the quantum error-correction era opening up at neutral-atom companies and of Rigetti’s 108-qubit Cepheus modular system, both of which depend on exactly the kind of fast classical decoding that Ising is designed to accelerate.