NollaMD's reported visual-medicine performance supports a specific proposition: a customised model can perform strongly on a defined task against the particular alternatives tested. It does not establish that the model outperforms every current general-purpose AI system, diagnoses patients better than doctors or validates Nolla Health's autonomous-prescribing pathway.
The relevant development report is TrainLoop's account dated 9 April 2026. This article examines that report as developer-published research and separates its findings from the additional evidence a clinical service would need. It is published by iatroX and includes iatroX's reference and learning tools in that wider discussion, not in a purported head-to-head benchmark. TrainLoop's NollaMD report.
What the developers actually report
TrainLoop describes adapting MedGemma 27B for visual medicine. In its experiment, supervised fine-tuning and preference-based training produced disappointing image-understanding results. A reinforcement-learning approach using multiple-choice tasks worked better. The report describes a held-out evaluation and lower costs than the frontier-model versions tested. These are results from that development programme, not universal conclusions about training methods.
The report also includes an important qualification: the team observed particularly strong performance on rare conditions but relatively average, or sometimes weaker, performance on common ones. Its stated next step involves obtaining missing context through follow-up conversation. These qualifications belong beside the headline, not in a footnote. TrainLoop, 9 April 2026.
Those findings deserve attention without being inflated. They suggest that specialist training may change a model's strengths in ways a broad average does not reveal. They do not establish which exact model version or combination of systems operates every part of a later patient-facing product.
Image interpretation is one part of a clinical encounter
Visual recognition asks what patterns an image supports. Differential diagnosis asks which explanations remain plausible when those patterns are considered with relevant context. Treatment selection adds another question: what action, including no immediate treatment or further assessment, is appropriate for this person?
These tasks overlap, but success at one does not logically establish success at all of them. A model could identify a plausible condition from a high-quality photograph while still needing history that is unavailable in the image. It could also identify the right condition without knowing that a particular intervention is unsuitable.
Consider a fictional evaluation item containing a clear image and a fixed selection of possible diagnoses. Now remove the answer options, add an ambiguous history and ask whether the patient fits a remote-treatment pathway. The second task is not simply the first task with a longer answer. It changes the information requirements, the permitted uncertainty and the consequences of being wrong.
That is why the link between NollaMD's research and Nolla's care service should be treated as a question about system design, not assumed identity. A service contains intake, eligibility rules, clinical operations, prescribing and follow-up in addition to any image model.
What does beating frontier models mean?
A benchmark comparison has a task, a dataset, a scoring rule and a set of model versions. Its meaning is bounded by those choices. It should not be rewritten as a timeless claim about all general-purpose AI.
The clinically useful questions are specific. Were the images representative of intended use? Were related images or patients kept apart across training and evaluation? Did the scoring distinguish a correct leading diagnosis from a correct diagnosis somewhere in a longer list? Were uncertain outputs and unanswerable cases allowed?
These questions are not allegations of methodological problems in NollaMD. They identify information needed to interpret any visual-medicine comparison. A careful reader can take a positive result seriously while asking how much of it transfers to the intended service.
Cost comparisons need the same discipline. A lower inference cost on the tested task may be useful, but it is not the cost of delivering care. Image collection, clinical review, escalation, support, medicine fulfilment and follow-up remain separate activities. A model comparison does not settle those operational costs.
Why common versus rare conditions matters
The prevalence of conditions in a test set can differ from their prevalence in a service. A model's average may therefore answer a different question from the one a patient or clinician actually cares about.
Imagine two fictional test sets. One deliberately includes many unusual visual presentations to probe breadth. The other reflects the cases arriving at a narrowly defined acne service. A gain on the first does not guarantee a gain on the second. Both evaluations can be useful, but they serve different purposes.
Similarly, a service must recognise people who should not enter its treatment pathway. A test containing only eligible cases cannot establish how reliably the system rejects a similar-looking but unsuitable presentation. The relevant evaluation needs both the expected cases and the boundary cases.
TrainLoop's common-condition caveat is therefore a reason to request task-specific evidence. It is not proof that Nolla's acne service performs poorly. Treating it as such would make the same mistake as treating a favourable aggregate result as proof of prescribing safety: transferring a finding beyond the question it measured.
Evaluate uncertainty and input quality, not just the answer
A proposed external evaluation should include poor lighting, incomplete views, irrelevant images, contradictory history and cases outside the intended scope. It should examine whether the model asks for clarification, declines to classify or confidently selects an unsupported option.
Correct abstention deserves its own analysis. A service that avoids every difficult case may be safe in one narrow sense but provide little useful access. A service that answers everything may look productive while concealing uncertainty. The balance should be assessed against appropriate clinical alternatives, not an invented error-free comparator.
Performance across skin tones also needs direct testing. A 2024 Nature Medicine study found that image-based AI assistance improved physicians' diagnostic performance while disparities persisted and, for generalists, increased. That was a different decision-support system, not a NollaMD trial; it illustrates why model performance and human-AI performance cannot be treated as identical.
This also means that adding a clinician does not end the evaluation. Reviewers need to recognise when the system is wrong, and the interface needs to make the relevant information available. The combination may perform differently from either component alone.
What evidence would connect the model to care?
The next evidential step is not merely another attractive benchmark chart. It is a description of the production workflow, the model's exact role and the decisions made by other components or clinicians.
A proposed service evaluation would then examine appropriate treatment, adverse events, successful escalation, patient-reported benefit and follow-up completeness. It would identify the model and pathway versions used, because changing either may change the result. It should also make clear whether evaluation was independent and whether the patients resembled the intended users.
No such head-to-head prescribing-outcome test has been run for this article. The distinction is especially important because Nolla's October 2026 Utah announcement describes a bounded protocol-driven service, not simply an unrestricted conversation with the research model.
A verdict by task, not a model league table
For a researcher developing a visual-medicine component, NollaMD's report provides a concrete specialist-training experiment to examine. For a clinician choosing a patient-care service, the benchmark is background information; the relevant evidence concerns the complete pathway. For someone needing broader evidence synthesis or an open-ended explanation, the model's visual benchmark does not establish which tool best fits that task.
A good learning experience is a further, distinct question. As described in iatroX's September 2026 product information, its Tutor starts from a learner's attempted question and probes the reasoning behind it. That design is about understanding a misconception, not maximising performance on a visual-diagnosis leaderboard. iatroX Tutor.
The useful conclusion is not that specialist models or generalists always win. It is that the task should determine the comparison, and the evidence should support the claim actually being made.
Frequently asked questions
What is NollaMD based on?
TrainLoop's report dated 9 April 2026 describes a customised MedGemma 27B model developed for visual medicine. That research description does not disclose the complete architecture of every Nolla patient-facing service.
Has NollaMD been independently shown to outperform doctors?
The material reviewed here does not establish that claim. Developer-reported results against selected AI models are not an independent comparison with doctors delivering complete patient care.
Does a diagnostic benchmark validate autonomous prescribing?
No, prescribing also depends on patient assessment, suitability, contraindications, operational safeguards and follow-up. Those parts of the service require their own evidence.
Explore reasoning-focused medical learning with iatroX Tutor →