RadKev decision model beats same-size LLM and hosted models on radiology tasks

[medRxiv] A medRxiv preprint reports that researchers fine-tuned the general-purpose decision models Kev-27B and Kev-9B on 36,109 records from public radiology datasets to produce RadKev-27B and RadKev-9B, then compared them with the Kev models, the hosted decision models Jev and OpenAI Decisions, and the LLM Qwen3.8-27B on 14,142 held-out questions across fifteen tasks. The prespecified primary outcome was task-averaged accuracy relative to Kev-27B. According to the authors, RadKev-27B exceeded Kev-27B by 4.9 percentage points (95% CI 3.8 to 6.2), with the largest gain in report error detection at 22.6 points; RadKev-9B exceeded Kev-9B by 6.9 points (5.5 to 8.4) and did not differ detectably from Kev-27B, a model three times its size. On 1,675 questions, RadKev-27B was more accurate than Jev (3.0 points, 1.7 to 4.4), OpenAI Decisions (4.8, 3.2 to 6.4), Qwen3.8-27B without reasoning (4.7, 3.4 to 6.0) and with reasoning (2.8 points, 1.0 to 4.6), at 132-fold lower latency (128 ms versus 16.9 s per question). At a 5% error rate, RadKev-27B could answer 90.5% of questions and Kev-27B 76.4%. The authors said the model can be deployed within an institution for categorical radiology decisions after validation on local data.