Nvidia has launched a new group of open AI models called Ising to help improve quantum computers.
The company announced the launch on Tuesday and said these models are built to handle two major problems in quantum computing today: tuning quantum processors properly and fixing errors quickly while the system is running.
According to Nvidia, these models are meant to support the bigger goal of fault-tolerant quantum computing.
Quantum computers rely on qubits, and those qubits are much more sensitive than the bits used in traditional computers.
As a result, mistakes still happen often, even in highly advanced quantum machines. That is why the technology is not yet ready for broad use in science or business.
Nvidia said today’s best quantum processors can still make about one error in every thousand operations.
For these systems to become truly useful at a large scale, that error rate may need to fall to around one in a trillion. The company says AI could play a major role in helping close that gap.
Two models for two major problems
Nvidia launched Ising with two purpose-built models. The first, Ising Calibration, is a vision-language model that reads results from quantum hardware experiments and recommends tuning steps.
It works a lot like a skilled technician, looking over the data and automatically making machine settings better.
Nvidia reported that it trained the model on real-world data of partners developing various quantum systems, such as superconducting qubits, quantum dots, trapped ions, and neutral atoms.
Ising Decoding is made to catch and fix errors while a quantum computer is working. Nvidia said it uses small AI models called 3D convolutional neural networks to do this job.
The company launched two versions. The Fast model has about 912,000 parameters and is built to run quickly on GPUs. The Accurate model has about 1.79 million parameters and is better at dealing with bigger and more difficult error patterns.
In Nvidia’s tests, the Fast version was 2.5 times faster and 1.11 times more accurate than the standard PyMatching decoder. The Accurate version was 2.25 times faster and 1.53 times more accurate.
Nvidia is releasing Ising as an open resource for developers. Users can download the model weights from Hugging Face and work with Nvidia tools like NIM and the NGC catalog to train and deploy the models.
They can also run and fine-tune everything on their own machines, which means sensitive data doesn’t have to leave their systems.
To measure progress, Nvidia and its partners created QCalEval, which the company called the first benchmark for AI-driven quantum calibration.
The test measures how well a model can read results, classify what happened, and suggest the next step.
Nvidia said Ising-Calibration-1 outperformed several leading models on that benchmark, including Gemini 3.1 Pro, Claude Opus 4.6, and GPT 5.4.
Could Nvidia’s Ising AI models threaten Bitcoin security?
Nvidia’s Ising AI models do not threaten Bitcoin. Ising is not a quantum computer. It is an AI system designed to help researchers calibrate and troubleshoot experimental quantum machines.
That efficiency could shorten development cycles and bring useful quantum applications closer to reality.
The tools run on standard GPUs and cannot break cryptographic codes. Bitcoin is encrypted with Elliptic Curve Digital Signature Algorithm (ECDSA), which would need a large, fault-tolerant quantum computer to execute Shor’s algorithm and compromise.
Experts say such a machine is still many years away, likely not before 2030. The timing of Nvidia launch also matters because quantum computing is starting to get more attention in the crypto world.
In a 31 March post, Google Research said future quantum computers may be able to break the encryption used by cryptocurrencies more easily than earlier estimates suggested.
Google said a machine with fewer than 500,000 physical qubits could do this within minutes under normal assumptions. It also warned that the crypto industry should start preparing stronger post-quantum security.
Another research update from the California Institute of Technology (Caltech) and Oratomic added to those concerns. The team said a fully realized fault-tolerant quantum computer may need only 10,000 to 20,000 qubits, far below earlier estimates that called for millions.
Caltech said that lower requirements could, in theory, make useful quantum machines possible by the end of the decade.