The question
Many radiology tasks ask for one answer from a known set: whether a report contains a finding, how urgently a study should be read, or which diagnosis best fits a case. A decision model returns a probability for each allowed answer instead of generating free text.
The project asked whether an open-weight decision model specialized for radiology could improve accuracy while remaining fast enough to run within an institution.
The approach
We fine-tuned Kev-27B and Kev-9B on 36,109 records from public radiology datasets, then evaluated the resulting RadKev models on 14,142 held-out questions across fifteen report-reading, diagnosis, classification, and knowledge tasks.
The comparison included the original Kev models, two hosted decision models, and a same-size language model with and without reasoning. The prespecified primary outcome was task-averaged accuracy relative to Kev-27B.
What the evaluation found
RadKev-27B exceeded Kev-27B by 4.9 percentage points in task-averaged accuracy. RadKev-9B exceeded Kev-9B by 6.9 points and did not differ detectably from Kev-27B, despite being one-third its size.
On the subset used for the language-model comparison, RadKev-27B was more accurate and returned an answer in 128 milliseconds on average, compared with 16.9 seconds for reasoning with Qwen3.8-27B. These are benchmark results, not evidence of clinical readiness. A hospital would still need to validate the model on its own data and workflow.
Open record
I am the submitting and first-listed author of the preprint with Ran Zhang. medRxiv posted version 1 on October 13, 2026; it has not been peer reviewed.
The training, evaluation, and statistical-analysis code is public. Model adapters and pointer heads are released for non-commercial research use under the terms stated in the repository and model cards.
Sources & further reading
Page last reviewed October 14, 2026.