# NVIDIA Ising

> Source: https://aiwiki.ai/wiki/nvidia_ising
> Summary: NVIDIA Ising is a family of open AI models from NVIDIA for operating quantum computers, covering two of the field's main engineering bottlenecks: quantum processor calibration and quantum error-correction decoding.
> Updated: 2026-08-07
> Fact-checked: 2026-08-07
> Categories: AI Models, NVIDIA, Open Source AI, Quantum Computing
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "NVIDIA Ising." aiwiki.ai, 7 Aug 2026. https://aiwiki.ai/wiki/nvidia_ising
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

| Field | Value |
|---|---|
| Developer | [NVIDIA](https://aiwiki.ai/wiki/nvidia) |
| Type | Open AI model family for quantum computing (calibration models and error-correction decoders) |
| Initial release | April 14, 2026 |
| Latest release | Ising Calibration 1.5 (July 2026) |
| Architectures | Vision language models (Qwen3.5-35B-A3B, Gemma 4 31B); 3D convolutional neural network decoders |
| Benchmark | QCalEval |
| Licenses | OpenMDW 1.1 (Calibration 1.5, ColorCode decoder); NVIDIA Open Model License (Calibration 1) |
| Website | developer.nvidia.com/ising |

**NVIDIA Ising** is a family of open AI models from [NVIDIA](https://aiwiki.ai/wiki/nvidia) for operating quantum computers, covering two of the field's main engineering bottlenecks: [quantum processor](https://aiwiki.ai/wiki/quantum_processor) calibration and quantum error-correction decoding. NVIDIA announced the family on April 14, 2026 as what it called "the world's first family of open source quantum AI models" [1]. Its calibration branch consists of [vision language models](https://aiwiki.ai/wiki/vision_language_model) that read plots produced by quantum calibration experiments and describe what they show, whether the experiment succeeded, and what to tune next; its decoding branch consists of [convolutional neural network](https://aiwiki.ai/wiki/convolutional_neural_network) pre-decoders for surface and color error-correcting codes. The most recent calibration release, Ising Calibration 1.5, is a 31-billion-parameter model built on [Gemma 4](https://aiwiki.ai/wiki/gemma_4) 31B, published in July 2026 under the [Linux Foundation](https://aiwiki.ai/wiki/linux_foundation)'s [OpenMDW](https://aiwiki.ai/wiki/openmdw) license in BF16 and [NVFP4](https://aiwiki.ai/wiki/nvfp4) versions [3][5][6].

## Background

Bringing a quantum processor into a usable state, and keeping it there, requires continuous calibration: qubit frequencies, gate parameters, and readout settings drift, and operators diagnose the drift by running experiments whose results are usually inspected as plots. NVIDIA's launch materials describe calibration and error correction as "two of the most critical challenges in building hybrid-quantum classical systems," and position AI models as a way to automate work that otherwise consumes days of expert time per bring-up [1]. The family is named after the Ising model, which the press release describes as "a landmark mathematical model that dramatically simplified the understanding of complex physical systems" [1].

The QCalEval paper that accompanies the calibration models frames the problem concretely: calibration plots are "the most universal human-readable representation" of quantum calibration data, yet before this work there was no systematic evaluation of how well vision language models interpret them [4].

## Launch

NVIDIA published the launch press release on April 14, 2026, alongside a special address from its Quantum Day event [1]. The initial family had two branches: Ising Calibration, a vision language model for interpreting quantum processor measurements, and Ising Decoding, two variants of a 3D convolutional neural network (one tuned for speed, one for accuracy) for real-time error-correction decoding. NVIDIA claimed the decoders were "up to 2.5x faster and 3x more accurate than pyMatching, the current open source industry standard," and that agent-driven calibration could cut calibration time "from days to hours" [1]. [Jensen Huang](https://aiwiki.ai/wiki/jensen_huang) said AI is "essential to making quantum computing practical" and described Ising as making AI "the control plane" of quantum machines, "transforming fragile qubits to scalable and reliable quantum-GPU systems" [1].

The press release listed early users of Ising Calibration including Atom Computing, Academia Sinica, EeroQ, Conductor Quantum, Fermi National Accelerator Laboratory, Harvard's John A. Paulson School of Engineering and Applied Sciences, Infleqtion, IonQ, IQM Quantum Computers, Lawrence Berkeley National Laboratory's Advanced Quantum Testbed, Q-CTRL, and the U.K. National Physical Laboratory. Ising Decoding deployments were listed at Cornell University, EdenCode, Infleqtion, IQM, Quantum Elements, Sandia National Laboratories, SEEQC, UC San Diego, UC Santa Barbara, the University of Chicago, the University of Southern California, and Yonsei University [1]. NVIDIA positioned Ising within its open model portfolio next to [Nemotron](https://aiwiki.ai/wiki/nemotron), [NVIDIA Cosmos](https://aiwiki.ai/wiki/nvidia_cosmos), [NVIDIA Alpamayo](https://aiwiki.ai/wiki/nvidia_alpamayo), [Isaac GR00T](https://aiwiki.ai/wiki/nvidia_isaac_gr00t), and BioNeMo, and said the family complements CUDA-Q and integrates with the NVQLink QPU-GPU interconnect [1][2].

## Model releases

| Model | Release | Architecture and base | License |
|---|---|---|---|
| Ising Calibration 1 (35B-A3B) | April 14, 2026 | [Mixture-of-experts](https://aiwiki.ai/wiki/mixture_of_experts) VLM on Qwen3.5-35B-A3B; ~35B parameters, 3B active per token | NVIDIA Open Model License [7] |
| Ising Decoder SurfaceCode 1 Fast / Accurate | April 14, 2026 | 3D CNN pre-decoders for surface codes | NVIDIA Open Model License; gated access [8] |
| Ising Decoder ColorCode 1 Fast | July 2026 | 3D CNN pre-decoder for triangular color codes | OpenMDW 1.1 [9] |
| Ising Calibration 1.5 (31B, BF16 and NVFP4) | July 2026 | Dense VLM on Gemma 4 31B; ~31B parameters | OpenMDW 1.1 [5][6] |

### Ising Calibration 1

The first calibration model is a fine-tuned derivative of Qwen3.5-35B-A3B, a [mixture-of-experts](https://aiwiki.ai/wiki/mixture_of_experts) model from the [Qwen](https://aiwiki.ai/wiki/qwen) family (roughly 35 billion total parameters, 3 billion active per token, 262,144-token context), trained with two-phase sequential supervised [fine-tuning](https://aiwiki.ai/wiki/fine_tuning) on 72.5K entries. NVIDIA's model card gives an April 14, 2026 release date, recommends a minimum of two L40S GPUs or one H100, and licenses the weights under the NVIDIA Open Model License, with Apache 2.0 applying to the Qwen base [7]. The QCalEval paper releases it as a "reference case study" and reports a 74.7 [zero-shot](https://aiwiki.ai/wiki/zero_shot_learning) average score on the benchmark [4].

### Ising Calibration 1.5

Ising Calibration 1.5 switched the base model: it is a dense multimodal VLM built on Google's Gemma 4 31B rather than a Qwen MoE [5]. The weights appeared on [Hugging Face](https://aiwiki.ai/wiki/hugging_face) on July 13, 2026, the [NIM](https://aiwiki.ai/wiki/nvidia_nim) container was published on NGC on July 23, and NVIDIA announced the release in a July 27 developer blog post by Tom Lubowe, the senior product manager for NVIDIA Ising, cuQuantum, and cuTENSOR, and Shuxiang Cao, an NVIDIA senior research scientist and the QCalEval paper's first author [3][5].

According to the blog post, the new model analyzes unfamiliar diagnostic results without prior training examples, uses examples from related experiments when available ([in-context learning](https://aiwiki.ai/wiki/in_context_learning)), and is 11.4% smaller at BF16 precision than its predecessor, easing local deployment of agentic calibration workflows [3]. The parameter counts on the model cards match that framing: roughly 31 billion versus the predecessor's roughly 35 billion [5][7]. For the first time the model also ships in an NVFP4-quantized version, which NVIDIA says runs on a single GPU, a consumer gaming card, or a [DGX Spark](https://aiwiki.ai/wiki/nvidia_dgx_spark) "with only a small cost in accuracy"; the full-precision model targets data center GPUs such as [Grace Blackwell](https://aiwiki.ai/wiki/nvidia_blackwell) and [Vera Rubin](https://aiwiki.ai/wiki/nvidia_vera_rubin) [3][6]. The blog also says tokens-per-second throughput on DGX Spark was optimized, with batching aimed at users running one agent across multiple experiments per qubit [3].

## Training data

NVIDIA's announcement says Ising Calibration 1.5 is trained on data generated from partner contributions across multiple qubit modalities, including superconducting qubits, quantum dots, ions, neutral atoms, and electrons on helium [3]. The model card describes the corpus in more detail: 72.5K supervised entries, split into 23.8K ICL-formatted entries with multi-image demonstrations and 48.7K zero-shot entries augmented using Qwen3.5-397B-A17B, with data collection and labeling both listed as synthetic, under a million training images, and under a billion text tokens [5].

One documented partner contribution comes from Northwestern University and Fermilab, whose NEXUS facility (107 meters underground at Fermilab) supplied superconducting qubit data from a month-long measurement campaign, including charge-jump measurements, used in training and in a charge tomography benchmark. Grace Bratrud, the NEXUS experimental lead, said the model "will be a great tool for identifying jumps in future datasets and could even enable real-time jump identification" [17].

## QCalEval benchmark

The models are evaluated with QCalEval, introduced in an April 28, 2026 arXiv paper ("QCalEval: Benchmarking Vision-Language Models for Quantum Calibration Plot Understanding") by 32 authors led by Cao, with co-authors including Alán Aspuru-Guzik and Krysta Svore [4]. The paper describes it as the first VLM benchmark for quantum calibration plots: 243 samples across 87 scenario types from 22 experiment families, spanning superconducting qubits and neutral atoms, with six question types (technical description, experimental conclusion, experimental significance, fit quality assessment, parameter extraction, and experiment success classification) evaluated in both zero-shot and in-context settings. The paper reports that the best general-purpose zero-shot model reached a mean score of 72.3, that many open-weight models degrade under multi-image in-context learning while frontier closed models improve substantially, and that supervised fine-tuning at the 9-billion-parameter scale improves zero-shot performance but does not close the in-context gap [4]. The dataset is published on Hugging Face under CC BY 4.0, with evaluation scripts in an Apache-licensed GitHub repository [10][11].

For Ising Calibration 1.5, NVIDIA reports the following QCalEval mean scores on the model card, averaged across GPT-5.4 and Gemini-3.1-Pro judges; the release candidate was evaluated on 243 zero-shot examples with 1,458 response slots and 236 ICL examples with 708 response slots [5].

| Setting | Ising Calibration 1.5 | Gemma-4-31B-IT (base model) |
|---|---:|---:|
| Zero-shot mean | 74.5 | 68.8 |
| MM-ICL mean | 81.2 | 81.2 |

NVIDIA's headline claims for the release are relative improvements, and the company's own materials state them inconsistently. The blog post's body says the model is "86.68% better than its predecessor" when using examples from related experiments, while the same post's figure caption puts the in-context improvement at 86.5%, the figure NVIDIA's Asia Pacific account repeated on X; the caption also says the model scores 10% better zero-shot on average than the next best open model of comparable size [3][12]. All of these describe improvement over Ising Calibration 1 or over rival models on QCalEval, not absolute accuracy. The blog says the model "outperforms all open models out of the box" and "remains competitive with leading closed models," naming Fable 5 and GPT 5.6 Sol; NVIDIA's Ising developer page, as accessed on August 8, 2026, separately claims Calibration 1.5 is 3.27% better than Gemini 3.1 Pro, 9.68% better than Claude Opus 4.6, and 14.5% better than GPT 5.4 [3][13].

## Error-correction decoders

The Ising Decoding models are AI pre-decoders: local, parallel networks that remove most physical errors before passing residual syndromes to a conventional global decoder such as PyMatching. An April 2026 paper on the surface-code models reports end-to-end decoding runtimes on the order of one microsecond per round at large code distances on GB300 GPUs, with logical error rates improved relative to global decoding alone, and a larger model variant outperforming correlated PyMatching up to distance 13 [14]. A July 2026 follow-up extends the approach to triangular color codes [15]. On its developer page NVIDIA quotes Decoder SurfaceCode 1 Fast at 2.5x faster latency and 1.1x higher accuracy than PyMatching at distance 13 and physical error rate 0.003, and Decoder ColorCode 1 at 7x faster latency and 347x higher accuracy than Chromobius at distance 31 [13]; the launch press release's broader "up to 2.5x faster and 3x more accurate" figure covers both speed- and accuracy-optimized variants [1].

Both decoder families sit behind a Hugging Face access request; the surface-code decoders carry the NVIDIA Open Model License, while the color-code decoder released in July 2026 switched to the OpenMDW 1.1 license [8][9].

## Deployment and licensing

Ising Calibration 1.5's checkpoints are published on Hugging Face in BF16 and NVFP4, both governed by the OpenMDW License Agreement version 1.1, a permissive machine-learning-model license stewarded by the Linux Foundation, with Apache 2.0 noted as additional information [5][6][16]. NVIDIA says the license offers QPU builders and operators "the flexibility to maintain data control and deploy anywhere" [3]. This is a change from Ising Calibration 1, which shipped under the NVIDIA Open Model License [7]. The model is also served as an NVIDIA NIM microservice with a [vLLM](https://aiwiki.ai/wiki/vllm) backend, hosted through NVIDIA Build, and documented in the NIM API reference [5][18].

For agentic use, NVIDIA publishes the Quantum-Calibration-Agent-Blueprint repository on GitHub (Apache 2.0, created March 31, 2026), a reference agent for discovering, executing, and analyzing calibration experiments; the blog describes it as a script for deploying an agentic workflow with Ising Calibration 1.5 using the [Nemo Agent Toolkit](https://aiwiki.ai/wiki/nemo_agent_toolkit), with support for other models including cloud APIs [3][19].

## Adoption and reception

Several partners published their own material on launch day. Q-CTRL described feeding Ising Calibration outputs into its Boulder Opal Scale Up autonomous calibration product, writing that "NVIDIA Ising can help us see clearly what the data is telling us, so our intelligent autonomy software can make the best decisions about how to maintain hardware without human intervention," and reported a demonstration on a 21-qubit QuantWare QPU in which the system characterized every component and calibrated single-qubit gates to 0.9988 median fidelity [20]. IQM announced it was integrating Ising-based AI agents into its existing calibration stack for enterprise systems [21]. The launch drew trade coverage from outlets including The Quantum Insider [2], and Quantum Zeitgeist covered the 1.5 release, noting the training span across six qubit modalities [22].

The launch press release cited analyst firm Resonance's projection that the [quantum computing](https://aiwiki.ai/wiki/quantum_computing) market will surpass $11 billion in 2030 as context for the release [1]. The model cards advise that outputs "should be validated by domain experts before acting on experimental conclusions" [5].

## See also

- [Quantum computing](https://aiwiki.ai/wiki/quantum_computing)
- [Quantum machine learning](https://aiwiki.ai/wiki/quantum_machine_learning)
- [Quantum processing unit](https://aiwiki.ai/wiki/quantum_processing_unit)
- [Vision language model](https://aiwiki.ai/wiki/vision_language_model)
- [OpenMDW](https://aiwiki.ai/wiki/openmdw)
- [NVIDIA DGX Spark](https://aiwiki.ai/wiki/nvidia_dgx_spark)
- [NVFP4](https://aiwiki.ai/wiki/nvfp4)

## References

1. "NVIDIA Launches Ising, the World's First Open AI Models to Accelerate the Path to Useful Quantum Computers." NVIDIA Newsroom, April 14, 2026. https://nvidianews.nvidia.com/news/nvidia-launches-ising-the-worlds-first-open-ai-models-to-accelerate-the-path-to-useful-quantum-computers
2. Swayne, Matt. "NVIDIA Launches Ising, the World's First Open AI Models to Accelerate The Path to Useful Quantum Computers." The Quantum Insider, April 14, 2026. https://thequantuminsider.com/2026/04/14/nvidia-launches-ising-the-worlds-first-open-ai-models-to-accelerate-the-path-to-useful-quantum-computers/
3. Lubowe, Tom, and Shuxiang Cao. "NVIDIA Ising Enables Fully Automated Quantum Computer Calibration with Enhanced In-Context Learning." NVIDIA Technical Blog, July 27, 2026. https://developer.nvidia.com/blog/nvidia-ising-enables-fully-automated-quantum-computer-calibration-with-enhanced-in-context-learning
4. Cao, Shuxiang, et al. "QCalEval: Benchmarking Vision-Language Models for Quantum Calibration Plot Understanding." arXiv:2604.25884, April 28, 2026. https://arxiv.org/abs/2604.25884
5. "NVIDIA-Ising-Calibration-1.5-31B-BF16." Model card, Hugging Face, July 2026. https://huggingface.co/nvidia/Ising-Calibration-1.5-31B-BF16
6. "NVIDIA-Ising-Calibration-1.5-31B-NVFP4." Model card, Hugging Face, July 2026. https://huggingface.co/nvidia/Ising-Calibration-1.5-31B-NVFP4
7. "NVIDIA-Ising-Calibration-1-35B-A3B." Model card, Hugging Face, April 2026. https://huggingface.co/nvidia/Ising-Calibration-1-35B-A3B
8. "Ising-Decoder-SurfaceCode-1-Fast." Hugging Face, April 2026. https://huggingface.co/nvidia/Ising-Decoder-SurfaceCode-1-Fast
9. "Ising-Decoder-ColorCode-1-Fast." Hugging Face, July 2026. https://huggingface.co/nvidia/Ising-Decoder-ColorCode-1-Fast
10. "QCalEval." Dataset, Hugging Face, April 2026. https://huggingface.co/datasets/nvidia/QCalEval
11. "NVIDIA/QCalEval." GitHub repository, April 2026. https://github.com/NVIDIA/QCalEval
12. NVIDIA Asia Pacific (@NVIDIAAP). "NVIDIA Ising Calibration 1.5, the latest from the NVIDIA Ising open model family, automates QPU calibration end-to-end..." X, August 3, 2026. https://x.com/NVIDIAAP/status/2084112023424528732
13. "NVIDIA Ising: AI Models & Framework for Quantum Computing." NVIDIA Developer, accessed August 8, 2026. https://developer.nvidia.com/ising
14. "Fast and accurate AI-based pre-decoders for surface codes." arXiv:2604.12841, April 14, 2026. https://arxiv.org/abs/2604.12841
15. "Fast and accurate AI-based pre-decoders for color codes." arXiv:2607.10058, July 11, 2026. https://arxiv.org/abs/2607.10058
16. "OpenMDW License Agreement, version 1.1." OpenMDW / Linux Foundation. https://openmdw.ai/license/1-1/
17. Sussman, Sara. "Northwestern and Fermilab Quantum Data Helps Build a New AI Benchmark for Quantum Calibration with NVIDIA Ising Open Models." Northwestern University Institute for Quantum Information Research and Engineering, April 14, 2026. https://quantum.northwestern.edu/news-and-stories/2026/northwestern-and-fermilab-quantum-data-helps-build-a-new-ai-benchmark-for-quantum-calibration-with-nvidia-ising-open-models.html
18. "nvidia / ising-calibration-1.5-31b." NVIDIA NIM API reference. https://docs.api.nvidia.com/nim/reference/nvidia-ising-calibration-1-5-31b
19. "NVIDIA/Quantum-Calibration-Agent-Blueprint." GitHub repository, March 2026. https://github.com/NVIDIA/Quantum-Calibration-Agent-Blueprint
20. Guilmart, James, and Alex Shih. "Scaling quantum autonomy with Q-CTRL's physics-informed AI and NVIDIA Ising." Q-CTRL, April 14, 2026. https://q-ctrl.com/blog/scaling-quantum-autonomy-with-nvidia-ising
21. "IQM Advances AI-Driven Agentic Calibration, Opening Quantum Computing to the Enterprise with NVIDIA Ising." IQM Quantum Computers, April 14, 2026. https://www.iqm.tech/press-releases/iqm-advances-ai-driven-agentic-calibration-opening-quantum-computing-to-the-enterprise-with-nvidia-ising
22. "NVIDIA's Quantum Calibration Model Works Across 6 Qubit Modalities." Quantum Zeitgeist, July 28, 2026. https://quantumzeitgeist.com/quantum-calibration-nvidias-works-across/