Exploratory comparison
All three models were evaluated on the same 500 ECGs from patients excluded from MCore fine-tuning. The ECGs were selected before any responses were generated.
The evaluation cohort contains 250 ECGs labeled NORM and 250 labeled non-NORM in PTB-XL. We randomly sampled eligible records using predefined quality and class balance criteria.
Table with columns: Model, Overall agreement, 250 normal ECGs, 250 abnormal ECGs| Model | Overall agreement | 250 normal ECGs | 250 abnormal ECGs |
|---|
| MCore ECG 0.8B | 367/500 · 73.4% | 189/250 · 75.6% | 178/250 · 71.2% |
| GPT-6.1 Terra · medium | 363/500 · 72.6% | 129/250 · 51.6% | 234/250 · 93.6% |
| GPT-6 Luna · medium | 354/500 · 70.8% | 133/250 · 53.2% | 221/250 · 88.4% |
We compared the overall normal or abnormal assessment extracted from each report with the corresponding NORM or non-NORM label. The metric evaluates that classification only. Indeterminate outputs and failed responses remain in the denominator and count as disagreements.
MCore had higher agreement on normal ECGs, while Terra and Luna had higher agreement on abnormal ECGs. The overall difference between MCore and Terra was four cases. The comparison provides descriptive results from this cohort.
All models received the same images and evaluation instructions, with inference settings specific to each model. We evaluated Terra and Luna through OpenRouter with medium reasoning effort. We selected the released MCore version based on its performance on this internal validation cohort.
Usage
Provide a legible 12-lead ECG image and ask the model to generate a report. To use the processor with Transformers, install a Torchvision version compatible with your PyTorch installation. The model returns an English report with an overall assessment, its findings, and any limitations it identifies.
The release includes FP16 Safetensors weights, configuration and processor files, and an F16 GGUF build with its matching vision projector. To use images with GGUF, load MCore-ECG-0.8B-F16.gguf together with mmproj-MCore-ECG-0.8B-F16.gguf in a runtime that supports Qwen3.5 multimodal inference.
Scope and limitations
This pilot is intended for research. The model can miss abnormalities or report findings or measurements unsupported by the tracing. Its suitability for diagnosis, triage, or clinical decision-making remains unestablished.
The evaluation covered rendered ECG images from PTB-XL. Performance on photographs, scans, other layouts, devices, and populations remains untested in this evaluation.
The fine-tuning targets are synthetic texts. Selected outputs received qualitative review. Comprehensive clinical review remains pending.
Sources and licenses
We release the weights under Apache 2.0, the same license as the base model, Qwen/Qwen3.5-0.8B. The license permits use, modification, and redistribution, including commercial use.
The ECGs come from PTB-XL 1.0.3 by Wagner et al., available under CC BY 4.0. We rendered the signals as 12-lead ECG images. The pilot adds synthetic text alongside the original source data.
See Wagner et al. (2020), PTB-XL, a large publicly available electrocardiography dataset, and the PTB-XL 1.0.3 dataset record.