NVIDIA research teams released “NVIDIA Ising Calibration 1.5” on July 27. This is a 31 billion-parameter visual language model designed specifically for diagnosing and tuning quantum processors, and it has been fully open-sourced. According to the QCalEval benchmark tests, the model’s performance on in-context learning (ICL) tests improved by 86.68% compared with the previous-generation model.
Ising Calibration 1.5 Performance Data
According to NVIDIA’s technical blog, the key performance figures for Ising Calibration 1.5 in the newly released QCalEval benchmark are as follows:
Zero-shot test: When assessing independent diagnostic results the model hasn’t seen before, the average performance is about 10% higher than other open-source models of the same scale.
ICL (in-context learning) test: When provided with relevant samples, performance saw a substantial 86.68% improvement over the previous-generation model. Its overall capability is now comparable to top closed-source models at the 1T-parameter level (such as Claude Fable 5, etc.).
NVFP4 quantized version lowers the deployment threshold: BF16 shrinks by 11.4%
According to NVIDIA’s technical briefing, the two key updates for Ising Calibration 1.5 are as follows:
NVFP4 quantized version: For the first time, NVIDIA provides an NVFP4 quantized version. With BF16 precision, the model size is reduced by 11.4%, enabling QPU diagnostic models that previously required massive compute to run smoothly now on a single GPU or on NVIDIA DGX Spark.
Supported quantum bit technologies: The training data for Ising Calibration 1.5 covers the following five mainstream quantum bit technologies:
· Superconducting qubits
· Quantum dots
· Trapped ions
· Neutral atoms
· Electrons on helium
Fully open-sourced and deployment methods
According to NVIDIA’s announcement, Ising Calibration 1.5’s complete resources (including parameter checkpoints, the quantized version, and the deployment blueprint) have been fully open-sourced under the Linux Foundation’s OpenMDW License. Developers can obtain and deploy it through the following channels: directly download the model weights on the Hugging Face platform, or quickly set up agentic calibration workflows via NVIDIA NIM containers and managed services.
FAQ
What are the main performance improvements of NVIDIA Ising Calibration 1.5?
Based on the QCalEval benchmark tests, Ising Calibration 1.5’s ICL performance improved by 86.68% over the previous-generation model. Its zero-shot performance is about 10% higher than other open-source models of the same scale, and its overall capability is already comparable to top closed-source models such as Claude Fable 5 at the 1T-parameter level.
How is the deployment threshold for Ising Calibration 1.5 lowered?
According to NVIDIA, the added NVFP4 quantized version reduces the BF16-precision model size by 11.4%. The model can now run on a single GPU or NVIDIA DGX Spark without requiring massive compute infrastructure.
How do I get Ising Calibration 1.5?
According to NVIDIA’s announcement, the model has been fully open-sourced via the Linux Foundation’s OpenMDW License. You can obtain the model weights on the Hugging Face platform, or deploy it via NVIDIA NIM containers and managed services.
