NVIDIA has launched Ising, the world's first family of open AI models designed to build fault-tolerant quantum systems. Quantum processors face a fundamental challenge: inherent noise leads to frequent errors, with current best processors erring once per thousand operations. To become useful for enterprise and scientific tasks, error rates must drop to one in a trillion. Ising leverages AI to address this gap through two primary domains: Ising Calibration and Ising Decoding. Both aim to reduce noise and correct errors in real time, a computationally intensive process required for scaling to millions of qubits. Ising Calibration utilizes Vision Language Models (VLMs) to analyze quantum experiment outputs and actively tune processors. The flagship model, Ising-Calibration-1, was trained on data from various qubit technologies, including superconducting and ion-based systems. It outperforms state-of-the-art models like Gemini 3.1 Pro and GPT-5.4 on QCalEval, the first benchmark for agentic quantum calibration. This model automates calibration workflows with minimal human oversight, integrating with tools like the NVIDIA NeMo Agent Toolkit. The second domain, Ising Decoding, focuses on quantum error correction. Using a specialized training framework based on NVIDIA cuQuantum and PyTorch, users can train 3D Convolutional Neural Network (CNN) decoders. These pre-decoders accelerate logical error correction by handling localized syndrome errors before global decoding. NVIDIA offers two base models to balance speed and accuracy. The Ising-Decoder-SurfaceCode-1-Fast model, with approximately 912,000 parameters, offers rapid inference suitable for low-latency requirements. Conversely, the Ising-Decoder-SurfaceCode-1-Accurate model features more layers and parameters to correct larger error chains, providing superior logical error rate improvements at the cost of runtime. Performance testing demonstrates significant gains. The accurate model combined with PyMatching achieved a 2.25x speedup and a 1.53x improvement in logical error rate compared to traditional methods. By running on NVIDIA GB300 GPUs, the system achieved round-trip latencies as low as 2.33 microseconds per round. These capabilities support scalable error correction strategies, including lattice surgery, which is essential for building large-scale quantum computers. The Ising family is fully open source. Weights, training frameworks, datasets, and benchmarks are available for developers to customize models for specific quantum processing unit (QPU) noise characteristics. NVIDIA provides pre-trained models on Hugging Face and via NVIDIA NIM and Build. The company also released a complete training framework enabling on-the-fly synthetic data generation and recipes for fine-tuning and quantization. By maintaining proprietary QPU data on-site while utilizing open AI workflows, Ising provides a flexible path toward Quantum-GPU Supercomputers capable of solving complex, real-world problems. This launch marks a critical step in bridging the gap between noisy intermediate-scale quantum devices and practical, fault-tolerant applications.