MedGemma 3 27B: Medical Foundation Model

• MedGemma 3 27B is a 27-billion parameter medical foundation model family derived from Gemma 3, available in both text-only and multimodal variants.
• Its training leverages post-training distillation and reinforcement learning on specialized medical QA and clinical datasets to enhance structured reasoning and extraction accuracy.
• Despite strong performance in medical QA and multilingual exam benchmarks, the model shows prompt sensitivity and hallucination risks, necessitating schema-guided clinical deployments.

MedGemma 3 27B is the 27-billion-parameter member of the MedGemma family, a collection of medical foundation models based on Gemma 3. In the literature, the designation appears in several closely related forms—MedGemma-27B, medgemma-27b-text-it, and, in some later work, “MedGemma 3 27B”. The technical report distinguishes a main 27B text-only model from a separately released 27B multimodal variant, while subsequent studies use the 27B model as a medical QA system, a local clinical NLP component, a hallucination-analysis target, and a benchmark baseline for outcome-grounded inpatient reasoning (Sellergren et al., 7 Jul 2025, Jonker et al., 5 May 2026).

1. Naming, scope, and model variants

The MedGemma literature uses overlapping but not identical naming conventions. The core technical report introduces MedGemma as “a collection of medical vision-language foundation models based on Gemma 3 4B and 27B,” but also states that the main MedGemma 27B in the report is a text-only model and that a distinct MedGemma 27B multimodal variant was also trained and released. Later benchmarking papers refer to medgemma-27b-text-it for the text instruction-tuned release, and some low-resource clinical QA work explicitly uses the phrase “MedGemma 3 27B” to denote the same 27B-scale Gemma-3-based medical model family (Sellergren et al., 7 Jul 2025, Carrillo-Larco et al., 15 Sep 2025, Jonker et al., 5 May 2026).

Name in the literature	Meaning in context	Modality
MedGemma 27B	Main 27B model in the technical report	Text-only
MedGemma 27B multimodal	Separate released variant with preliminary evaluation	Image + text
medgemma-27b-text-it	Instruction-tuned 27B text variant used in downstream evaluations	Text-only
MedGemma 3 27B	Later naming for the 27B Gemma-3-based medical model	Paper-dependent

This distinction is material. Results on medgemma-27b-text-it in medical examinations, EHR QA, and interpretability studies do not automatically transfer to the 27B multimodal model, and vice versa. A plausible implication is that “MedGemma 3 27B” functions more as a family-level identifier in later application papers than as a single invariant checkpoint name.

2. Architecture and training methodology

For the main 27B text model, medical specialization is introduced entirely in post-training rather than domain-adaptive pretraining at the 27B stage. The report states that MedGemma 27B is built on the Gemma 3 architecture and infrastructure, retains the base model’s general capabilities, and acquires medical expertise through distillation from a large instruction-tuned teacher and reinforcement learning on medical text datasets. The post-training corpora listed for the 27B text model include MedQA (9,275), MedMCQA (182,806), AfriMed-QA (1,003), MedExpQA (434), PubMedQA (1,000), LiveQA (634), HealthSearchQA (3,375), and 200,000 synthetic medical questions (Sellergren et al., 7 Jul 2025).

The multimodal pathway uses MedSigLIP, a medically tuned vision encoder derived from SigLIP-400M. The report states that MedSigLIP was fine-tuned on >33M medical image–text pairs, with medical data mixed at 2% during encoder enhancement, after which the multimodal decoder was re-adapted with medical image–text data mixed at 10% for ≈5 epochs. Multimodal post-training used RL rather than SFT, which the authors state generalized better. The Gemma image pathway uses 896×896 inputs with pixel values normalized to $[-1,1]$, while a standalone 448×448 MedSigLIP version is also released for efficiency (Sellergren et al., 7 Jul 2025).

The report also specifies a 128k-token context length and a SentencePiece tokenizer with 262,000 entries. Its stated language-modeling objective is

$$L_{LM} = -\sum_t \log p_\theta(w_t \mid w_{<t}, v),$$

where $v$ denotes visual context when present. For the vision encoder enhancement, it gives the SigLIP-style alignment objective

$$L_{SigLIP} = -\frac{1}{N} \sum_{i=1}^{N} \left[ \log \sigma(\alpha \cdot s_{ii}) + \sum_{j \neq i} \log \sigma(-\alpha \cdot s_{ij}) \right], \quad s_{ij} = f(x_i)^\top g(y_j).$$

These details place MedGemma 3 27B at the intersection of long-context medical language modeling and medical VLM development, but they also underscore a persistent bifurcation in the literature between the text-only 27B and the multimodal 27B (Sellergren et al., 7 Jul 2025).

3. Benchmark performance and core capability profile

The technical report presents MedGemma 27B as a strong text-only medical model. On the report’s medical QA benchmarks, MedGemma 27B scores 89.8 on MedQA (4-option) with test-time scaling, 74.2 on MedMCQA, 76.8 on PubMedQA, and 84.0 on AfriMed-QA; on MedXpertQA, it records 25.7 versus 15.7 for Gemma 3 27B. In the report’s agentic evaluation, it scores 56.2 on AgentClinic-MedQA versus 50.7 for Gemma 3 27B, and 46.0 on AgentClinic-MIMIC-IV versus 35.2, yielding a stated +10.8 point out-of-distribution gain on the latter. For EHRQA, the text-only 27B improves from 86.3 before RL fine-tuning to 93.6 after RL, reducing error from 13.7% to 6.4% (Sellergren et al., 7 Jul 2025).

The preliminary multimodal 27B appendix reports MIMIC-CXR macro-F1 90.0, CheXpert macro-F1 49.9, SLAKE tokenized F1 70.0, VQA-RAD tokenized F1 46.7, and PathMCQA accuracy 71.6. Because the report labels these results as preliminary and places them in an appendix specific to the multimodal release, they should be read as early characterization rather than as the canonical definition of MedGemma 3 27B performance (Sellergren et al., 7 Jul 2025).

Later task-specific evaluations refine the capability profile. On PeruMedQA, a Spanish MCQA benchmark of 8,380 Peruvian medical exam questions, medgemma-27b-text-it is the top model across years and specialties, with six exam slices above 90%: Psychiatry 2025: 94.00%, Pathology 2024: 92.00%, Test B 2020: 91.11%, Test B 2018: 91.00%, Pediatrics 2019: 91.00%, and Test A 2024: 91.00%. Its invalid answer rate is 0.02% (2/8,380), and 2.69% (n=226) of all questions are answered correctly only by this model (Carrillo-Larco et al., 15 Sep 2025).

These results suggest a model whose strongest documented text performance lies in structured medical QA, multilingual examination settings, and instruction-following-heavy tasks. That pattern is consistent with the report’s post-training recipe, which emphasizes distillation on medical QA corpora rather than full-domain pretraining at 27B.

4. Downstream clinical uses and prompt-based adaptation

MedGemma 3 27B has been deployed primarily through prompting, structured output constraints, and lightweight post-processing rather than through repeated full-model retraining. In heterogeneous EHR prescription extraction for medications for opioid use disorder, a system customizing open-source LLMs “including Llama, Qwen, Gemma, and MedGemma” extracts a unified set of prescription attributes and computes MOUD days per patient. On prescription-level data from five clinics, MedGemma-27B attains 93.1% coverage and 92.2% exact-match accuracy, trailing Qwen2.5-32B only narrowly. The reported workflow uses fixed JSON schemas, lightweight normalization, and cross-field consistency checks, and is framed as removing brittle, site-specific ETL while supporting local, privacy-preserving deployment (Fei et al., 23 Oct 2025).

In ArchEHR-QA 2026, MedGemma 3 27B is evaluated in a strict low-resource, no-weight-update setting for grounded EHR question answering. On the development set, it reaches up to 88.0 on the Subtask 1 composite metric under task decomposition and few-shot prompting, approximately the mid-30s on the patient-friendly generation composite for Subtask 3, and an average Micro F1 of 81.0 on evidence citation alignment in Subtask 4. The paper emphasizes that these results were obtained with prompt engineering alone, using temperature = 0.0 and top-p = 0.95, and that open-weights deployment is important under GDPR/HIPAA constraints; the open-source models under ~70B parameters, including MedGemma 3 27B, were run locally on up to two NVIDIA L4 (24GB) GPUs (Jonker et al., 5 May 2026).

The multimodal 27B variant also appears in inference-time optimization work. In PneumoniaMNIST, prompt-only MAP-Elites evolution is applied to MedGemma-27B multimodal without fine-tuning the model weights. At 224×224, the reported test accuracy improves from 83.33 ± 0.00 to 84.46 ± 0.14, and macro-F1 from 82.96 ± 0.00 to 83.88 ± 0.15. The gains are larger at lower resolutions, where the base prompt is weaker. The paper attributes the improvement to evolved finding-oriented rules, calibrated thresholds, and strict JSON-output preservation rather than to superficial prompt rewording (Sviridov et al., 5 Jun 2026).

Across these studies, MedGemma 3 27B is typically embedded in pipelines that impose schema control, decomposition, or evolutionary search over prompts. This suggests that its practical identity in the literature is not only that of a base model, but also that of a model whose clinical utility is often mediated by prompt programming and structured interfaces.

5. Robustness, failure modes, and internal representations

Several papers characterize MedGemma-27B as strong but nontrivially brittle. In prompt-sensitivity experiments on MedMCQA and PubMedQA, MedGemma exhibits family-level degradation under common prompting heuristics: Chain-of-Thought decreases accuracy by 5.7% relative to direct answering, few-shot examples degrade performance by 11.9%, and shuffling answer options changes predictions 59.1% of the time, with accuracy drops up to 27.4 percentage points. For the 27B model specifically, cloze scoring reaches 64.5% on MedMCQA with position bias 0.054, outperforming all prompting strategies tested. The cloze score is defined as

$$s(o) = \sum_{t \in tokens(o)} \log p(t \mid context), \qquad \hat y = \arg\max_o s(o).$$

The authors interpret this as evidence that the model “knows” more than its generated text shows (Sadanandan et al., 26 Mar 2026).

Bias-oriented evaluation yields a related picture. In a counterfactual cultural-cues benchmark built from MedQA items, MedGemma-27B has an Original option-only accuracy of 72.00% and Neutral accuracy of 72.67%. Under option-only prompting, the largest average degradation occurs when Identifier + Context cues co-occur: −3.11 percentage points from Original and −3.78 from Neutral. Under brief-explanation prompting, the baseline is 75.51% for both Original and Neutral, and the corresponding Identifier + Context degradation shrinks to −0.57 percentage points. Yet more than half of culturally grounded explanations still end in incorrect answers: 62.67% for identifier-only, 54.35% for context-only, and 61.18% for identifier-plus-context (Rezaei et al., 27 Jan 2026).

Hallucination analysis in GI endoscopy VQA adds another axis of robustness. On Gut-VLM, MedGemma-27B has hallucination prevalence 61.1% across 4,392 QA pairs. Among detection methods, the white-box ReXTrust achieves AUC 90.39 and AUPRC 92.50, outperforming the best gray-box alternative, MaxProb, at AUC 80.93 and AUPRC 81.99. The study also identifies confident confabulation—hallucinated answers with high inter-sample agreement and/or high token-level confidence—as a systemic failure mode, occurring in 11.6% of MedGemma-27B’s hallucinated samples (Lawal et al., 23 Jun 2026).

Interpretability work further shows that Gemma3-27B and MedGemma-27B have activations that collapse at intermediate layers but recover by the final layers. For MedGemma-27B, the best linear recoverability of age is $R^2 = 0.98 \pm 0.01$ at layer 24, disease-class clustering is strongest—though still weak—at layer 61, and drug clustering peaks around layers 44–45. The same study reports strong late-layer saliency in Gemma/MedGemma and severe degradation under layer lesioning, implying a relatively distributed and fragile knowledge organization rather than a clean early-layer concentration (Marinescu et al., 13 Oct 2025).

Taken together, these results argue against treating MedGemma 3 27B as a uniformly robust medical oracle. The evidence instead points to a model with high latent medical competence, measurable prompt sensitivity, bias susceptibility under non-decisive cues, and failure modes that become more legible when hidden-state access is available.

6. Clinical reasoning, deployment context, and open questions

Outcome-grounded inpatient reasoning exposes an additional limitation of the off-the-shelf 27B model. In CLR-POMDP, a benchmark that frames inpatient reasoning as a partially observable decision problem, MedGemma-27B scores 66.66% aggregate, whereas CLR-voyance-8B reaches 84.91% after rubric-based GRPO post-training. The same paper also uses MedGemma-27B-text-it as an LLM-as-a-judge, where it achieves 77.3% accuracy and $\kappa = 0.421$ zero-shot, and 77.7% accuracy with $\kappa = 0.399$ under 5-shot ICL. This suggests that MedGemma-27B can align reasonably with clinician-authored rubrics as a grader, even while underperforming a smaller model specialized for outcome-aware reasoning (Nagar et al., 10 May 2026).

The deployment picture is correspondingly mixed. On one hand, MedGemma is openly released through Health AI Developer Foundations, with tutorials and model weights, and its open-weights status makes it attractive for local clinical deployment where privacy or licensing constrain API use. On the other hand, the technical report explicitly cautions that the models are not clinically validated, that the 27B multimodal evaluation remains preliminary, and that many current benchmarks may suffer from saturation or leakage, motivating higher-fidelity and application-level validation (Sellergren et al., 7 Jul 2025).

A concise synthesis is therefore possible. MedGemma 3 27B is best understood as a high-capacity, Gemma-3-based medical foundation model family centered on a strong 27B text model and accompanied by a separate 27B multimodal release. It performs strongly on medical QA, multilingual exam benchmarks, structured extraction, and low-resource prompt-based EHR tasks; it can be improved further through RL fine-tuning, prompt evolution, or schema-constrained orchestration. At the same time, the literature documents substantial prompt sensitivity, hallucination-related vulnerabilities, culture-cue instability, and incomplete transfer to outcome-grounded inpatient reasoning. This suggests that MedGemma 3 27B is most reliable not as a standalone endpoint, but as a component within tightly specified clinical pipelines, explicit evaluation protocols, and domain-specific adaptation regimes (Sellergren et al., 7 Jul 2025, Nagar et al., 10 May 2026).