NVIDIA Ising

全球首个面向量子计算校准与纠错的开源AI模型系列

large language model

https://github.com/NVIDIA/ising

Visit Website

In-depth Report

NVIDIA Ising (Ising) is the world's first open source AI model series for quantum computing scenarios released by NVIDIA at the GTC conference on April 14, 2026 (World Quantum Day). This model addresses the two core problems of long time-consuming calibration of quantum processors and low efficiency of quantum error correction, and provides a complete open source model, tools and tutorials from quantum device calibration to quantum error correction. Ising compresses calibration time from days to hours, increases error correction speed by 2.5 times, and increases accuracy by 3 times, marking a key transition for quantum computing from laboratory to industrialization.
Ising is independently developed by NVIDIA and belongs to the NVIDIA quantum computing product line. On April 14, 2026, Huang Renxun officially released this model series at the World Quantum Day (GTC Conference). NVIDIA said these AI models will help research institutions and enterprises build better quantum computers that can run practical applications at scale. As an open source project, Ising is released on GitHub under the Apache 2.0 license and has currently received 52 stars and 18 forks. The core value of Ising is to provide complete open source models, tools and tutorials from quantum device calibration to quantum error correction. Jensen Huang, founder and CEO of NVIDIA, said: "AI is crucial to the practical realization of quantum computing. Through the Ising model, AI will become the control plane of quantum machines - the operating system - transforming fragile qubits into scalable, highly reliable quantum GPU systems."
The Ising model series includes a variety of models with different functions, covering key aspects of the entire quantum computing process. In terms of quantum device calibration, Ising-Calibration-1-35B-A3B is a 35B parameter visual language model (VLM) specifically used for quantum device calibration. This model can automatically complete the calibration work that originally took several days, greatly improving the startup efficiency of quantum computers. In terms of quantum error correction, Ising-Decoder-SurfaceCode-1-Accurate is a high-precision AI pre-decoder for surface code quantum error correction, containing 1.79M parameters. There is also a fast version of Ising-Decoder-SurfaceCode-1-Fast, which contains 0.91M parameters and provides a flexible choice between accuracy and speed. These decoders increase error correction speed by 2.5 times and accuracy by 3 times. Ising also provides a complete open source tool chain, including AI decoding training framework (Ising-Decoding), quantum computer calibration agent workflow blueprint (Quantum Calibration Agent Blueprint), VLM quantum calibration experiment analysis capability open benchmark (QCalEval) and other code repositories. Detailed tutorial documents (Cookbooks) are also provided, covering the training of AI pre-decoder, optimized inference and quantization, as well as the introduction to quantum computer calibration agent workflow. Ising supports multiple deployment methods. Inference microservices can be deployed directly through Nvidia Inference Microservices (NIMs), called through the Build.nvidia.com API endpoint, or downloaded from HuggingFace.
The Ising model series adopts a completely open source and free business model. All models, tools and tutorial resources are free and open to scientific research institutions, enterprises and developers around the world. This reflects NVIDIA's strategic layout to promote the development of the quantum computing ecosystem, lower the threshold of quantum computing through open source, and accelerate the practical process of quantum computing. Although there is no direct business model, the release of Ising helps NVIDIA establish technological leadership in the field of quantum computing, form synergies with NVIDIA's GPU computing platform, and lay the foundation for possible future enterprise-level services.
Since Ising was just released on April 14, 2026, the number of public user comments is currently limited. Judging from feedback from the technical community, scientific researchers have shown high interest in Ising. The academic community's evaluation of Ising is relatively positive. This project lowers the threshold for quantum computing research, allowing researchers without a professional background in machine learning to use AI for quantum computing calibration and error correction. The pre-trained model provided by Ising-Decoder can be directly deployed and used, greatly reducing development costs. The developer community recognizes open source protocols and documentation for completeness. The Apache 2.0 license provides greater freedom of use, and the Cookbooks tutorial covers the complete process from training to deployment.
The quantum computing industry responded positively to Ising's release. Some evaluations believe that Ising has overcome the two core problems of long time-consuming calibration of quantum processors and low quantum error correction efficiency, and is an important breakthrough in the practical application of quantum computing. From an industry perspective, Ising marks a key transition for quantum computing from the laboratory to industrialization. By applying AI technology to the core bottleneck of quantum computing, NVIDIA demonstrates its strategic layout in the field of quantum computing. The release of Ising also reflects the trend of accelerated integration of AI and quantum computing. Error correction is considered the key to practical quantum computing, and Ising is reshaping this field.
So far, there has been no major controversy surrounding the Ising project itself. However, as a cutting-edge technology product, users need to pay attention to the following risk factors. In terms of technology maturity risks, Ising is an emerging project, and the actual deployment effect will take time to be verified. Quantum computing hardware itself is still under development, and it remains to be seen whether breakthroughs at the software level can actually translate into actual performance improvements. In addition, competition in the field of quantum computing is becoming increasingly fierce, and whether Ising can maintain its technological leadership requires continued attention. Competitors such as Google and IBM are also increasing investment in quantum computing AI.
Ising is suitable for the following user groups: Quantum computing researchers can use Ising models to conduct quantum error correction research without building an AI decoder from scratch; quantum computing engineers can use calibration tools to automate workflows and improve equipment utilization efficiency; AI developers can explore the intersection of quantum computing and AI and participate in open source contributions; university scientific research teams can carry out quantum computing teaching and experiments based on Ising. For traditional IT developers, Ising requires a certain background knowledge of quantum computing to use it effectively. Ordinary users are advised to learn the basic concepts of quantum computing first.
NVIDIA Ising is a milestone product in the field of quantum computing AI, promoting the practical process of quantum computing through open source. Its calibration and error correction capabilities are expected to accelerate the popularization of quantum computers, but the actual industrial effect still needs to be verified.

User Reviews

MAgon

—

英伟达这波把量子AI做成了开源，格局是真大。

Deborah.Cox_99

—

它那个35B模型喂了7万2千多条专家数据做两阶段SFT，所以输出校准建议时确实有点老研究员的味道，不是那种泛泛的通用VLM。

BrooklynnScott

—

470倍加速有点离谱，先观望。

d7qb8fnx2

—

我用两块L40S把Ising校准跑起来了，原本要数天的人工校准流程，现在几个小时就出结果，最关键是它直接看示波器波形和频谱图就给校准建议，不像以前还要自己盯着数据猜。

BCruzII46

—

说点真实的。Ising Decoding比pyMatching快2.5倍、准3倍这个数据，在纠错这种分秒必争的场景确实有用，毕竟量子态相干时间就那么长。但问题是它现在只支持超导量子比特路线，离子阱、光量子、中性原子全都没影儿，等于把生态锁在NVIDIA想推的那条路上，做拓扑纠错研究的人暂时用不上。

Christine.Collins0078

—

重点不是这个模型多强，是英伟达在量子硬件还没成熟前，先把软件层和操作系统的位置占住了。等真用量子计算机跑业务的时候，研究机构早就把CUDA-Q和Ising这套用顺手了。

JButler_594

—

等AMD版，手头全是MI300X跑不了有点憋屈。

小鱼_24

—

35B参数的VLM塞进校准流程，262K上下文确实能看长实验日志。

Margaret.Edwards_X609

—

我们组拿Ising Calibration做了个域外实验，把它用在投资组合的QAOA电路质量评估上，在L4上跑出373倍GPU加速，Sharpe比朴素经典选法中位数高3.1%，还在Rigetti Cepheus-108Q上花1.44美元验了一把。必须诚实说这不算量子优势，只是混合工作流验证，但Ising确实能抓出QAOA收敛差的问题。NVIDIA把模型开源后才有人这么快玩出花，这点得服。

Kyle_Reed007501

—

部署门槛比想象低，一块H100或者两块L40S就能跑，对实验室来说不算夸张。但是4-bit量化走的是bitsandbytes的CUDA专用路径，AMD和苹果芯片官方明确不支持，社区以后会不会移植另说。

SSimmons21

—

QCalEval上它比Claude Opus 4.6高快十个点，量子校准这块闭源模型真被它超了。

暖阳164

—

免费的，先star了再说。

oejdj3l

—

冷静看，那篇470倍加速的论文是英伟达自己报的，arxiv 2604.02179，独立复现还没出来。对比IBM Osprey说精度高22%、耗时才32%，两边硬件原理、编码效率、成本结构根本不一样，这更像给自家软件优先路线占位的市场话术，不能直接当经典模拟优于量子计算的证据。真要信得等第三方bench。

JoshuaHorton

—

Ising纠错解码用的是3D CNN，天生适合处理带时空相关性的拓扑纠错数据，Fast版91万参数、Accurate版179万参数，两个变体一个拼速度一个拼精度，设计挺务实的。

TaylorKim

—

有没有人试过把它接到自己的量子实验平台？NeMo Agent那套自动化校准链路好接吗，文档看着还行但没敢真上产线。

Christina.Cook168

—

校准从「算出来」变成「看出来」，这个思路很英伟达。VLM直接读测量图生成结构化技术文本，再喂给智能体驱动连续校准，把高度依赖专家经验的数天活儿压到几小时，省的不是算力是人力。

ERobinson520

—

量子界的CUDA时刻？

CMorgan32

—

Apache2.0开源这点对中小企业太关键了，权重、训练代码、框架接口全放出来，允许商用和修改。现在量子方向的工具大多要么闭源要么学术玩具，英伟达这套直接给生产级的东西，等于把开发者社区往CUDA-Q生态里拽。短期看是卖GPU的钩子，长期看量子软件栈的标准可能真被它定了。

SawyerSchmidt

—

提醒一句，Ising Decoding完全基于合成数据训练，靠在模拟环境里加去极化噪声造样本。真实硬件噪声分布跟模拟不一定对齐，上线前最好拿自家QPU的数据再验一遍再信它的3倍精度。

Rachel.Butler_7

—

四月十四号发布的，名字取自伊辛模型，寓意用AI把量子系统的复杂问题「简化」掉，挺巧思的。

Frank.Stephens168

—

看到NVIDIA发布Ising模型我真的惊了！量子计算终于有救了，校准时间从几天缩短到几小时，这波操作太给力了。

天涯_14

—

作为一个量子计算方向的研究生，Ising开源真的是及时雨。之前每次实验前都要花好几天校准设备，现在有了AI模型终于可以省点时间了。

邓雪艳

—

NVIDIA从CUDA-Q到NVQLink再到Ising，这布局也太清晰了吧！感觉量子计算生态越来越完整了。

Emma.YoungII

—

校准速度提升这么多，纠错速度2.5倍，准确率3倍，这是真的强。唯一担心的就是实际部署效果能不能达到论文说的水平。

Raymond.Davis_99

—

看了下GitHub上的代码和文档，Apache 2.0协议真的很友好。相比某些厂商藏着掖着的态度，NVIDIA这波大气了。

yellowmouse651

—

TomHardware和TechPowerUp都报道了，看来业界关注度很高啊！

PlutoProtocolCampbell

—

期待效果。之前用传统方法做量子纠错累死累活，AI介入后希望能轻松点。

Jean.Lewis_2022

—

免费开源真的香！之前用别的方案光是校准服务就要花不少钱，现在可以直接白票。

唐玉

—

Ising-Decoder的Fast和Accurate两个版本可以满足不同需求，很灵活。点赞。

DomingoOrtiz

—

全球首个开源量子AI模型，这个title就很霸气。Google和IBM要紧张了吧。

NVIDIA Ising

In-depth Report

User Reviews

Similar Products