NVIDIA Ising: The World's First Open-Source AI Models for Quantum Computing

Quantum computing's biggest bottleneck isn't building qubits — it's keeping them stable. Calibration takes days, and quantum error correction (QEC) demands real-time processing that existing tools struggle to deliver. On April 14, 2026, NVIDIA announced Ising, the world's first open-source AI model family designed to solve both problems.

What Is Ising?

Named after the mathematical Ising Model that simplified complex physical systems, NVIDIA's Ising is a family of AI models targeting two core quantum computing challenges:

Calibration — keeping quantum processors tuned and accurate
Quantum Error Correction (QEC) — fixing errors in real-time during computation

Jensen Huang's framing: "AI becomes the operating system of quantum machines — transforming fragile qubits into scalable, reliable quantum-GPU systems."

Ising Calibration: A 35B Vision-Language Model

The calibration model is a 35 billion parameter VLM that interprets quantum processor measurements and automates continuous calibration. The result: what used to take days now takes hours.

Benchmark Results (QCalEval)

NVIDIA co-developed QCalEval, the world's first agent-based quantum computer calibration benchmark. Ising Calibration's performance vs. general-purpose models:

Model	Ising Advantage
Gemini 3.1 Pro	+3.27%
Claude Opus 4.6	+9.68%
GPT 5.4	+14.5%

Domain specialization clearly matters. A 35B model purpose-built for quantum calibration outperforms much larger general-purpose models on this specific task.

Ising Decoding: Real-Time QEC with Tiny Models

The decoding component uses 3D convolutional neural networks in two variants:

Speed-optimized (0.9M parameters)
Accuracy-optimized (1.8M parameters)

Compared to pyMatching (current open-source standard):

2.5x faster
3x more accurate

The small model sizes are the key insight here. At under 2M parameters, these models can run directly in quantum processor control loops for real-time error correction.

Fully Open Source

NVIDIA released everything:

Model weights
Training framework
Training data
Benchmarks (QCalEval)
Training recipes

Available on Hugging Face, GitHub, and build.nvidia.com. The NVIDIA Open Model License allows fine-tuning with proprietary QPU data while keeping it local. NIM microservices enable minimal-setup fine-tuning.

Platform Integration

Ising integrates with NVIDIA's quantum computing stack:

CUDA-Q: Hybrid quantum-classical computing platform
NVQLink: QPU-GPU hardware interconnect for real-time control
cuQuantum / cuStabilizer: Synthetic data generation and training

Global Adoption

24 institutions are already adopting Ising:

Calibration: IonQ, IQM, Harvard SEAS, Infleqtion, Q-CTRL, Fermi National Lab, and more.

Decoding: Cornell, UC Santa Barbara, Sandia National Labs, University of Chicago, Yonsei University, and more.

Why This Matters for Developers

The quantum computing market is projected to exceed $11 billion by 2030, but that growth depends on solving QEC and scalability. NVIDIA is positioning AI as the control plane — the operating system layer that makes quantum hardware usable.

For developers interested in quantum computing, Ising provides:

Pre-trained models you can fine-tune on your own QPU data
Benchmarks (QCalEval) to evaluate your own approaches
Integration path via CUDA-Q for hybrid quantum-classical workflows

Ising joins NVIDIA's growing open model portfolio: Nemotron (agents), Cosmos (physics AI), Alpamayo (autonomous driving), Isaac GR00T (robotics), and BioNeMo (biomedical research).

Source: NVIDIA Official Announcement