NVIDIA Ising: Open AI Models for Fault-Tolerant Quantum

On April 14, 2026 — World Quantum Day — NVIDIA shipped Ising, an open-source AI model family aimed at two of the hardest problems between today's noisy quantum processors and tomorrow's useful ones: real-time error-correction decoding and processor calibration.¹² NVIDIA describes Ising as the world's first family of open AI models for quantum computing.¹ The release pairs a 35B-parameter vision-language model trained on quantum experiment plots with a pair of compact 3D CNN decoders that NVIDIA benchmarks as up to 2.5x faster and 3x more accurate than pyMatching, the most widely used open-source decoder for surface-code error correction.³⁴

TL;DR

Ising is a two-part family. Ising Calibration is a Mixture-of-Experts vision-language model — Ising-Calibration-1-35B-A3B, 35B total parameters with ~3B active per token across 256 experts (8 routed per token) — built on top of Qwen3.5-35B-A3B with an integrated vision encoder for experiment plots.⁵ On NVIDIA's own QCalEval benchmark (243 entries across 87 scenarios from 22 experiment families) it scores 3.27% above Gemini 3.1 Pro, 9.68% above Claude Opus 4.6, and 14.5% above GPT-5.4.⁶ Ising Decoding is a 3D convolutional neural network that ships in two variants — Fast (0.9M parameters) and Accurate (1.8M parameters) — for real-time surface-code error correction.⁴ At distance d=13 with physical error rate p=0.003, the Fast variant is 2.5x faster and 1.1x more accurate than pyMatching; the Accurate variant is 2.3x faster and 1.5x more accurate. NVIDIA also reports the decoder requires roughly 10x less training data than alternative approaches.⁴⁷ Weights are released under the NVIDIA Open Model License; training frameworks are Apache 2.0 on GitHub; both checkpoints are on Hugging Face.⁸

What You'll Learn

• The two model components inside Ising and what each one solves
• Why April 14 — and why "Ising" is the name
• The Calibration MoE architecture and what QCalEval actually measures
• The Decoding 3D-CNN variants and the pyMatching baseline
• How Ising plugs into CUDA-Q QEC and NVQLink for real-time loops
• Licensing, availability, and which labs are already adopting it

Why April 14 — And Why "Ising"

Two pieces of branding context matter for understanding the launch. First, April 14 is World Quantum Day, an initiative quantum scientists started in 2021 to mark Planck's constant — 4.1356677×10⁻¹⁵ eV·s — whose first three rounded digits give the calendar date.⁹ NVIDIA's announcement timed itself to that observance.

Second, the name "Ising" points to the early-twentieth-century Ising model, a lattice model of interacting spins that became one of the foundational simplifications of complex many-body physics.¹² The metaphor lines up with what these AI models do: take a noisy, high-dimensional measurement signal off a quantum chip, and collapse it into something tractable.

Ising Calibration: A 35B Vision-Language Model for Qubit Plots

Calibrating a quantum processor is, in practice, hours of staring at plots. Engineers run an experiment, look at a heatmap or a Rabi oscillation, and decide what to tune next. The pipeline is visual, multi-step, and currently human-bound. Ising Calibration is NVIDIA's attempt to automate that loop.

The model is a Mixture-of-Experts vision-language model. The architecture pairs an integrated vision encoder for experimental plot images with the Qwen3.5-35B-A3B MoE language model, giving roughly 35 billion total parameters with ~3 billion active per token — 256 experts in the MoE layer, 8 routed and 1 shared activated for each token.⁵ The model accepts a plot image plus a structured natural-language question and emits structured technical text.

NVIDIA evaluated it on a benchmark it published alongside the model, QCalEval, which it describes as the first vision-language benchmark for quantum calibration plots.¹⁰

On QCalEval, Ising Calibration 1 scores higher than the closed frontier models that came up in the same comparison: NVIDIA reports it is 3.27% better on average than Gemini 3.1 Pro, 9.68% better than Claude Opus 4.6, and 14.5% better than GPT-5.4.⁶ These are NVIDIA-reported numbers on a NVIDIA-published benchmark, so the right framing is "the calibration-tuned model beats general-purpose VLMs at the calibration task it was trained for" rather than "this is a more capable general VLM than GPT-5.4."

Ising Decoding: Two 3D CNNs Against the pyMatching Baseline

If calibration is the slow human-in-the-loop problem, decoding is the fast machine-only one. A real fault-tolerant quantum computer running surface-code error correction has to read syndrome measurements off the chip and infer the most likely set of physical errors — within microseconds, before the next round of gates breaks coherence. Today most groups use pyMatching, an open-source minimum-weight perfect matching decoder, as the standard baseline.⁴

NVIDIA's Ising Decoding takes a different approach. The decoder is a 3D convolutional neural network that ingests a stacked syndrome volume (two spatial dimensions of the surface code, plus rounds of measurement on the third axis) and emits a correction. Two variants ship at launch:

Source: NVIDIA Ising release.⁴⁷

Two structural things matter beyond the headline numbers. First, because the decoder is a 3D CNN, the same trained weights run on different surface-code distances and round counts at inference time — the model is simply applied to a different decoding volume, with no retraining required to change distance or rounds.⁴ Independent third-party work has noted that maintaining the error-correction advantage at substantially larger code distances may require scaling the decoder size proportionally, but small models work across a useful range out of the box.¹¹ Second, NVIDIA reports the decoder requires roughly 10x less training data than alternative approaches at comparable accuracy.⁷

For deployment, the trained model exports to ONNX, then runs through NVIDIA TensorRT with FP8 or FP16 inference on GPU. Optional INT8 or FP8 post-training quantization is supported.¹²

How It Plugs Into CUDA-Q and NVQLink

Ising is not a standalone artifact — it is the AI half of a quantum-classical loop NVIDIA has been assembling since late 2025.

NVQLink, announced October 28, 2025 at GTC Washington D.C., is the GPU-to-QPU interconnect: an open system architecture that uses RDMA over Converged Ethernet (RoCE) on standard NVIDIA networking hardware to deliver deterministic communication between quantum control electronics and GPU servers, with a maximum round-trip latency of 3.96 microseconds.¹³ It is the "wire" between the quantum chip and the AI decoder.

CUDA-Q QEC 0.6, released the same day as Ising, is the software side. It introduces two new real-time-capable decoder pipelines over NVQLink: RelayBP, a GPU-accelerated belief-propagation decoder for quantum LDPC codes, and the NVIDIA Ising pre-decoder combined with PyMatching as a global decoder for the surface code.¹²¹⁴ The pre-decoder pattern — neural net does the fast, high-accuracy first pass, classical algorithm cleans up residual cases — is how Ising drops into a working error-correction loop without replacing the existing pipeline wholesale.

Why this matters: the entire round-trip latency, from qubit measurement to corrective gate, has to fit inside the coherence window of the underlying qubits. For superconducting qubits operating on tens-of-nanoseconds gate times and surface-code cycle times of around 1 microsecond, decoders need to keep up in real time — the closer to the cycle time, the better.¹⁵¹⁶ A decoder-only speedup of 2.5x is therefore not a benchmark trophy; it is what closes the gap between offline analysis and a working real-time QEC loop.

The Wider Quantum Decoder Race

Ising lands in a busy field. In March 2026, QpiAI demonstrated a union-find decoder on a 64-qubit Kaveri processor that completed each error-correction cycle in roughly 1.5 microseconds, fast enough to run inside the coherence window of superconducting qubits.¹⁷ Google's Willow processor — disclosed in late 2024 and refined since — runs a real-time decoder on a 101-qubit distance-7 surface code, achieving 0.143% ± 0.003% error per cycle and beating its best physical qubit's lifetime by a factor of 2.4 ± 0.3.¹⁶ AlphaQubit, a Google-DeepMind transformer-based decoder published in Nature in November 2024, was demonstrated on Google's Sycamore processor at distance-3 and distance-5 surface codes, with simulated runs out to distance 11.¹⁸ A separate convolutional neural-network decoder (Gicev et al., 2023, Quantum) has been demonstrated on simulated surface codes with code distances exceeding 1000 — over 4 million physical qubits — the largest ML-based surface-code decoder demonstration to date.¹⁹

These results are not all directly comparable — Willow's 0.143% per cycle is a full-stack hardware-plus-decoder logical error rate, QpiAI's 1.5 µs is a hardware-decoder cycle time on a single Kaveri device, and NVIDIA's 2.5x speedup is a decoder-only benchmark against pyMatching. What Ising adds to the picture is a different kind of contribution: open weights, an open training framework, an open benchmark, and a deployment pipeline that any research group with NVIDIA GPUs can adopt.

Licensing and Availability

Source: NVIDIA Ising release notes and Hugging Face model cards.⁸¹²

The weights/code split — model weights under NVIDIA's Open Model License, training framework under Apache 2.0 — is the same pattern NVIDIA used for Nemotron and recent NIM-distributed models. It gives QPU operators latitude to keep proprietary calibration data on-prem while still letting them fine-tune the published checkpoints for their specific hardware.

Early Adopters

NVIDIA listed seven early-adopting institutions in the launch press release: Academia Sinica, Fermi National Accelerator Laboratory, Harvard John A. Paulson School of Engineering and Applied Sciences, Infleqtion, IQM Quantum Computers, Lawrence Berkeley National Laboratory's Advanced Quantum Testbed, and the U.K. National Physical Laboratory (NPL).¹² The mix — three national labs, two academic groups, and two commercial QPU companies — spans both hardware modalities the QCalEval benchmark covers, with Infleqtion on neutral atoms and IQM on superconducting qubits. Subsequent NVIDIA materials have listed additional adopters including IonQ, Atom Computing, EeroQ, Q-CTRL, and Conductor Quantum.⁴

Bottom Line

NVIDIA Ising is the first time a major hyperscaler has shipped open weights, an open training framework, and an open benchmark targeted specifically at the calibration and error-correction subproblems of quantum computing — paired with a real-time deployment pipeline (NVQLink + CUDA-Q QEC) that lets the AI half of the loop actually meet the microsecond budget the physics demands. Whether the 2.5x latency and 3x accuracy numbers hold up under independent replication on third-party QPUs is the question to watch over the next two quarters. But the architectural bet — that the path to useful quantum computers runs through AI-accelerated classical co-processors, not just better qubits — is now openly testable by anyone with an NVIDIA GPU and a quantum testbed.

For broader context on AI in scientific computing, see our earlier coverage of thermodynamic computing as a post-transistor frontier and the DeepSeek V4 open-source frontier release.