Nvidia has released Ising, a new family of AI models for quantum processor calibration and quantum error-correction decoding, and said the models are available under its Open Model License on GitHub, Hugging Face and build.nvidia.com.
The company said the release is meant to help researchers and enterprises build quantum processors capable of running useful applications, with Ising designed to work alongside CUDA-Q, Nvidia’s software platform for hybrid quantum-classical computing, and NVQLink, its link between quantum processors and GPU systems.
Why calibration and decoding are the core problem
Calibration and decoding represent the primary classical bottlenecks in achieving fault-tolerant quantum computing. Calibration is the work of keeping quantum hardware tuned and in spec, while decoding is the classical software step that interprets error data fast enough for quantum error correction.
In a technical post accompanying the launch, Nvidia wrote that the best quantum processors still make an error roughly once in every thousand operations and that useful systems will require error rates closer to one in a trillion.
Fault tolerance means building a system that can keep running useful workloads even though qubits are noisy and errors are frequent.
Riverlane, which focuses on quantum error correction, separately says today’s best quantum computers can perform only a few thousand reliable quantum operations before failure and will need to scale to millions and then trillions of operations for broader practical use.
What Ising actually consists of
According to Nvidia, Ising Calibration is a 35-billion-parameter vision-language model that can analyze quantum calibration experiment output and be used in agentic workflows with minimal human oversight.
It said Ising Decoding consists of two AI models, one tuned for speed and one for accuracy, for real-time quantum error correction. Nvidia’s model card for Ising Calibration says the model is built on Qwen3.5-35B-A3B, with about 35 billion total parameters and roughly 3 billion active per token.
For decoding, Nvidia compared Ising against pyMatching, which it described as the current open-source industry standard. In its technical blog, the company said the fast model plus pyMatching was 2.5 times faster than pyMatching alone and 1.11 times more accurate in one benchmark setup.
It said the larger model plus pyMatching was 2.25 times faster and 1.53 times more accurate under the same benchmark conditions, and reported a threefold improvement in logical error rate in another benchmark regime.
A new benchmark because none existed
Calibration claims come with a new benchmark rather than an established industry standard. Nvidia and its research collaborators said they created QCalEval because no standard benchmark existed for quantum calibration models, and described it as the first vision-language benchmark for quantum calibration plots.
Nvidia says QCalEval tests whether a model can interpret calibration results, classify outcomes, assess fit quality and recommend the next step.
The research page says QCalEval contains 243 samples across 87 scenario types from 22 experiment families, while Nvidia’s technical post said Ising Calibration outperformed Gemini 3.1 Pro, Claude Opus 4.6 and GPT 5.4 on that benchmark.
Infrastructure Nvidia had already been building
Nvidia introduced NVQLink in October 2025 as an interconnect for coupling quantum processors to GPU systems, and said at the time that the architecture was meant to support control, calibration and quantum error correction.
Last month, IonQ said it had announced a memorandum of understanding with South Korea’s KISTI to pursue integration of IonQ hardware with high-performance computing infrastructure using NVQLink.
Nvidia listed Atom Computing, IonQ, IQM Quantum Computers, Q-CTRL and several national labs and universities among early users of the Ising models.
Decoding and calibration as operational barriers
Infleqtion, in a same-day post about its own neutral-atom roadmap, separately said it is integrating Ising Decoding and described decoding as a central bottleneck because every round of quantum error correction generates information that must be processed quickly enough to keep up with the hardware.
The criticality of real-time decoding was detailed in a December 2024 Nature paper from Google Quantum AI and collaborators, which said a fault-tolerant quantum computer requires a classical co-processor that can decode errors in real time and warned that if the system cannot process those error signals, known as syndromes, fast enough, backlogs can sharply increase computation time.
Calibration, too, is now being discussed as an operational barrier rather than just a research chore. In an April 8 announcement with Equal1, Q-CTRL said a primary barrier to broader adoption is the complexity of booting and maintaining quantum hardware, which it said is still typically handled manually by teams of PhD-level experts.
In a separate April 14 post tied to Nvidia Ising, Q-CTRL said the sector’s focus is shifting from raising qubit counts to building the operational intelligence needed to control larger systems.