NVIDIA Ising Calibration is an open supply imaginative and prescient language mannequin (VLM) designed to interpret diagnostic outputs from quantum processors and decide how they need to be tuned to proceed working.
This submit introduces the most recent mannequin launch, NVIDIA Ising Calibration 1.5, which advances AI-based QPU calibration by analyzing unfamiliar diagnostic outcomes with out prior coaching examples. Ising Calibration 1.5 additionally makes use of examples from associated experiments when obtainable and is 11.4% smaller at BF16 precision. This eases the deployment of agentic calibration workflows immediately in native lab environments.
For the primary time, the mannequin can also be obtainable in an NVFP4-quantized model, enabling deployment on a single GPU or an NVIDIA DGX Spark—comparable with main closed fashions corresponding to Fable 5 and GPT 5.6 Sol.
How is the Ising Calibration 1.5 mannequin educated?
The Ising Calibration 1.5 mannequin is educated on knowledge generated from companion contributions throughout a number of qubit modalities, together with superconducting qubits, quantum dots, ions, impartial atoms, electrons on Helium, and others specializing in calibration and management.
How is Ising Calibration 1.5 efficiency evaluated?
Efficiency of Ising Calibration 1.5 is evaluated utilizing the QCalEval benchmark, which measures a mannequin’s capacity to interpret experimental outcomes, classify outcomes, consider significance, assess match high quality and key options, and suggest subsequent steps. For extra particulars on the benchmark, mannequin structure, and analysis outcomes, see QCalEval: Benchmarking Imaginative and prescient-Language Fashions for Quantum Calibration Plot Understanding.
The analysis covers each zero-shot and in-context studying (ICL). Zero-shot reasoning analyzes outcomes independently, whereas ICL evaluates ends in the context of associated samples. Each are essential for constructing brokers that may automate QPU bring-up and retune operations.
On the QCalEval benchmark, Ising Calibration 1.5 reveals robust efficiency when analyzing diagnostic outcomes with out prior examples. It’s now additionally 86.68% higher than its predecessor when utilizing examples from associated experiments. It outperforms comparable open fashions and stays aggressive with main closed fashions.
Ising Calibration 1.5 advances AI and quantum computing calibration by outperforming all open fashions out of the field and on the QCalEval benchmark. It’s aggressive with state-of-the-art closed or 1T+ parameter fashions.
The 31-billion-parameter VLM is fitted to knowledge heart GPUs corresponding to NVIDIA Grace Blackwell and NVIDIA Vera Rubin. It additionally ships with a quantized model to NVFP4 the place customers can run this mannequin on a client gaming card or NVIDIA DGX Spark with solely a small value in accuracy.
The tokens per second (TPS) efficiency on DGX Spark has additionally been optimized, enabling good throughput regionally for a fraction of the price.
To be taught extra about deploying Ising Calibration 1.5 with an agent, take a look at the NVIDIA/Quantum-Calibration-Agent-Blueprint GitHub repo.
Get began with NVIDIA Ising open assets
The NVIDIA Ising mannequin household is absolutely open. Weights, knowledge, benchmarks, and recipes are supplied so you may modify, deploy, and fine-tune your personal fashions and variants on your particular QPUs.
Mannequin weights
Full-parameter checkpoints for Ising Calibration 1.5 can be found on Hugging Face:
Ising Calibration 1.5 can also be obtainable as an NVIDIA NIM and hosted by way of NVIDIA Construct. The OpenMDW License from Linux Basis presents QPU builders and operators with the pliability to keep up knowledge management and deploy anyplace.
Deployment recipes
A ready-to-use agent harness blueprint is on the market with assist for this mannequin and others, together with massive mannequin cloud APIs.
Quantum calibration agent blueprint is a script for deploying an agentic workflow utilizing Ising Calibration 1.5 with the NVIDIA Nemo Agent Toolkit to rapidly arrange quantum calibration experiment automation.
Open datasets and QCalEval benchmark
Ising Calibration 1.5 is constructed on actual QPU knowledge supplied by companions and collaborators. A semantic quantum calibration benchmark has additionally been launched to judge mannequin effectiveness for this job.
