NVIDIA-Ising-Calibration-1-35B-A3B
Model Summary
| |
|---|
| Total Parameters | ~35B total, 3B active per token (MoE sparse activation) |
| Architecture | Mixture-of-Experts Vision-Language Model (MoE VLM) |
| Base Model | Qwen3.5-35B-A3B (256 experts, 8 active per token) |
| Context Length | 262,144 tokens |
| Precision | BF16 |
| Input | Image (PNG, JPEG) + Text |
| Output | Text (technical analysis, conclusions, parameter extraction) |
| Training | Two-phase sequential SFT (72.5K entries) |
| Minimum GPU | 2x NVIDIA L40S (48GB) or 1x H100 (80GB) |
| Release Date | April 14, 2026 |
| License | NVIDIA Open Model License |
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
Use of the model is governed by the
NVIDIA Open Model License.
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. Use
temperature=0 and
max_tokens=16384 for optimal results.
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
Reference(s)
Benchmarks
| Question Type | Ising Cal 1 | Qwen3.5-35B base |
|---|
| Q1 Technical Description | 87.8 | 86.8 |
| Q2 Experimental Conclusion | 67.1 | 39.9 |
| Q3 Experimental Significance | 64.7 | 45.7 |
| Q4 Fit Quality Assessment | 90.5 | 52.7 |
| Q5 Parameter Extraction | 62.5 | 57.8 |
| Q6 Experiment Success | 75.3 | 50.6 |
| Overall | 74.7 | 55.5 |
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. Default inference settings use temperature=0 and max_tokens=16384. Context length: 262,144 tokens.
Output
Output Type(s): Text
Output Format(s):
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):
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
Recommended Inference Settings
| Parameter | Value |
|---|
temperature | 0 |
max_tokens | 16384 |
top_p | 1.0 |
Training and Evaluation Datasets
Training
Data Modality
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
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:
Labeling Method by dataset:
Properties: Curated quantum calibration experiments with ground-truth labels derived from simulation parameters.
Inference
Acceleration Engine: vLLM with FlashAttention, BF16 precision
Test Hardware:
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
1@misc{cao2026qcaleval,
2 title = {QCalEval: Benchmarking Vision-Language Models for Quantum Calibration Plot Understanding},
3 author = {Cao, Shuxiang and Zhang, Zijian and others},
4 year = {2026},
5 url = {https://research.nvidia.com/publication/2026-04_qcaleval-benchmarking-vision-language-models-quantum-calibration-plot},
6}
