NVIDIA-Ising-Calibration-1-35B-A3B

Model Summary

Description

NVIDIA-Ising-Calibration-1 analyzes quantum computing calibration experiment plots and generates structured technical text across six analysis question categories. NVIDIA-Ising-Calibration-1 was developed by NVIDIA as a quantum calibration vision-language model built on Qwen3.5-35B-A3B. This model is ready for commercial/non-commercial use.

Governing Terms

The Ising-Calibration-1-35B-A3B model is governed by the NVIDIA Open Model License Agreement. ADDITIONAL INFORMATION: For Qwen3.5-35B-A3B Apache License, Version 2.0.

Deployment Geography: Global

Quick Start

This model is a fine-tuned derivative of Qwen3.5-35B-A3B. Follow the Qwen3.5-35B-A3B serving guide for deployment with vLLM, replacing the model path with nvidia/NVIDIA-Ising-Calibration-1-35B-A3B. Suggested inference settings: temperature=0.2, max_tokens=16384.

Use Case

Quantum computing researchers, calibration engineers, and developers can use this model to analyze experiment plot images and generate technical descriptions, experimental conclusions, significance assessments, fit quality evaluations, parameter extractions, and experiment success classifications. Model outputs should be validated by domain experts before acting on experimental conclusions.

Release Date

Hugging Face 04/14/2026 via https://huggingface.co/nvidia/NVIDIA-Ising-Calibration-1-35B-A3B

Reference(s)

• Qwen3.5
• QCalEval Benchmark

Benchmarks

Evaluated on the QCalEval Benchmark: 243 entries across 87 scenario types from 22 experiment families, spanning superconducting qubits and neutral atoms. Scores are averaged across GPT-5.4 and Gemini-3.1-Pro judges.

Model Architecture

Architecture Type: Mixture-of-Experts Vision-Language Model (MoE VLM)

Network Architecture: Integrated vision encoder for experiment plot images combined with the Qwen3.5-35B-A3B MoE language model for autoregressive text generation.

This model was developed based on: Qwen3.5-35B-A3B

Number of model parameters: ~35B total parameters, ~3B active per token (256 experts, 8 active)

Training Methodology

Phase 1 — ICL-formatted SFT

• 23.8K in-context learning formatted entries
• Teaches the model to process multi-image demonstrations
• Learning rate: 1e-5, 1 epoch

Phase 2 — Zero-shot SFT

• 48.7K zero-shot entries, LLM-augmented via Qwen3.5-397B-A17B
• Strengthens single-plot understanding across all question types
• Learning rate: 5e-6, 1 epoch

Total training data: 72.5K entries

Input

Input Type(s): Image, Text

Input Format(s):

• Image: PNG, JPEG
• Text: String

Input Parameters:

• Image: Two-Dimensional (2D)
• Text: One-Dimensional (1D)

Other Properties Related to Input: Single-image or multi-image quantum calibration experiment plots with text prompts delivered through an OpenAI-compatible API. Suggested inference settings: temperature=0.2 and max_tokens=16384. Context length: 262,144 tokens.

Output

Output Type(s): Text

Output Format(s):

• Text: String

Output Parameters:

• Text: One-Dimensional (1D)

Other Properties Related to Output: Natural language technical analysis, experimental conclusions, significance assessments, fit quality evaluations, parameter extractions, and experiment success classifications.

Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA hardware and software frameworks, the model achieves faster inference times compared to CPU-only solutions.

Software Integration

Runtime Engine(s):

• vLLM with FlashAttention, BF16 serving precision

Supported Hardware Microarchitecture Compatibility:

• NVIDIA Blackwell (sm100, B200/B300/GB200)
• NVIDIA Hopper (sm90, H100/H200/GH200)

Supported Operating System(s):

• Linux (Ubuntu 22.04+)

The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks before deployment.

Model Version(s)

v1.0.0

Training and Evaluation Datasets

Training

Data Modality

• Image
• Text

Training Data Size

72.5K total entries (Phase 1: 23.8K ICL-formatted entries; Phase 2: 48.7K zero-shot entries).

Data Collection Method by dataset

• Synthetic (LLM-augmented via Qwen3.5-397B-A17B)

Labeling Method by dataset

• Synthetic

Properties: Synthetically generated quantum calibration experiment plots with paired analytical text.

Evaluation Dataset

Benchmark Score: QCalEval Benchmark zero-shot scores.

Description: QCalEval is a VLM benchmark for quantum calibration plots: 243 entries across 87 scenario types from 22 experiment families, spanning superconducting qubits and neutral atoms. It evaluates six question types: technical description (Q1), experimental conclusion (Q2), experimental significance (Q3), fit quality assessment (Q4), parameter extraction (Q5), and experiment success classification (Q6).

Data Collection Method by dataset:

• Synthetic

Labeling Method by dataset:

• Synthetic

Properties: Curated quantum calibration experiments with ground-truth labels derived from simulation parameters.

Inference

Acceleration Engine: vLLM with FlashAttention, BF16 precision Test Hardware:

• 2x NVIDIA L40S (48GB)

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and developers should ensure this model meets the requirements of their use case and addresses foreseeable misuse before deployment.

For more detailed information on ethical considerations for this model, please see the Model Card++ Bias, Explainability, Safety & Security, and Privacy Subcards.

Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here.

Citation

@misc{cao2026qcaleval,
  title  = {QCalEval: Benchmarking Vision-Language Models for Quantum Calibration Plot Understanding},
  author = {Cao, Shuxiang and Zhang, Zijian and others},
  year   = {2026},
  url    = {https://research.nvidia.com/publication/2026-04_qcaleval-benchmarking-vision-language-models-quantum-calibration-plot},
}