Models
Models & approach
How Meddox uses AI today — and what we are watching. Decision support only.
Not a medical device. Benchmarks are not clinical validation. We do not claim a “best medical AI.”
Meddox is a personal Akte with multi-model decision support. Scores on exams and leaderboards move quickly; they do not equal safe care. Always verify with original records and a clinician.
Last reviewed 21 October 2026 (Europe/Berlin).
Our approach (now)
Health content goes only to Vertex AI on the EU multi-region (default eu, never global). eu is not Frankfurt. Grok and the direct Anthropic API are for non-health calls unless a default-off flag is on. A pilot allowlist can turn a provider off. Published leaderboard scores are not a Meddox result.
Chat can add a specialty checklist from the diagnosis fields. MedGemma is not on the chat picker.
What the code calls
- Health calls use Vertex. Fast tasks ask for gemini-3.8-flash. Reasoning asks for gemini-4-argon. The default location is the EU multi-region eu (host aiplatform.eu.rep.googleapis.com). The global endpoint is refused. Model cards updated 21 October 2026 list those ids on eu and do not list them on europe-west3 (Frankfurt), which is why the default is eu. An explicit europe-west3 still fails closed and is not moved. London (europe-west2) is refused. eu is not a promise of processing in one country. An endpoint does not guarantee in-region processing. The direct Gemini API is not used for health data.
- Health data does not call api.anthropic.com. Claude on Vertex is off unless VERTEX_CLAUDE_HEALTH is on, and then only at location eu. The Sonnet 5.5 card (19 October 2026) names multi-region and global, not europe-west3. Non-health calls may still use claude-sonnet-5-5 on api.anthropic.com. claude-opus-5-5 is used only when ANTHROPIC_MODEL is set to that id on that non-health path.
- Grok is for non-health calls: grok-4.7 by default on api.x.ai, which sets no region. If that id is missing, the non-health path tries grok-4.6. xAI documents a United States regional endpoint. No EU endpoint was found. A health-data call throws unless XAI_HEALTH_ENABLED is on. That flag is off. This draft does not claim a DPA or zero-data retention.
- Speech to text of a record is health data. The route asks Vertex for gemini-3.5-transcribe. The card updated 21 October 2026 lists that id on global, us, and eu only, not europe-west3. The default location is eu, so the region check passes. An explicit europe-west3 fails closed and does not open xAI. grok-voice-transcribe-2.0 is used only if XAI_HEALTH_ENABLED is on. grok-voice-2.0 is not called.
- MedGemma stays off the chat picker and is not deployed. The imaging route does not call it. If the checkbox is on, the response records health_vertex_only and does not fetch. The stub remains in the repository. The model card says it is not clinical-grade. The default id is medgemma-1.5. MedGemma 1 27B weights exist. Meddox does not select them.
- A browser helper asks the health router before any fetch. The three direct hosts are refused for health content, so no request is sent. The keys are not NEXT_PUBLIC_ names and are not in the browser bundle. The helper is not a server route and is not behind the pilot allowlist. A local text generator continues. That is not a model call.
GeminiClaudeGrok
What exists (honest catalog)
Names and access paths change. This is orientation, not a shopping list, and it is not a list of models Meddox calls. Open-weight rows without a new card in this review are unchanged.
Closed API / research surfaces
| Name | Job | How you get it |
|---|---|---|
| Med-Gemini | Research name. No matching id on the Gemini API models page reviewed this date. | Not a Meddox route. |
| Med-PaLM 2 | Research model behind MedLM. Not called by Meddox. | Deprecated. Google Cloud says MedLM access ended on or after 29 September 2025. |
Open weights / self-host
| Name | Job | How you get it |
|---|---|---|
| MedGemma 1.5 4B | Open 4B multimodal weights. Model card 1.5.0 (13 January 2026), context at least 128K tokens. Google says it is not yet clinical-grade. Not a Meddox chat model. | Open weights or Model Garden. The imaging route does not call it. |
| Med42 v2 | Open clinical chat weights from M42 (8B and 70B). Text only. The v2 card lists an 8K token context. Not called by Meddox. | Hugging Face weights / self-host. Med42 v1 remains published and is older. |
| OpenBioLLM | Open biomedical language model | Open weights / self-host |
| BioMistral | Open biomedical LLM (Mistral lineage) | Open weights / self-host |
| Apollo | Open medical multilingual model family | Open weights / self-host |
| MeditronFO | Fully open medical training recipe / weights | Open recipe + weights — self-host |
Imaging specialists
| Name | Job | How you get it |
|---|---|---|
| MedSAM | 2D medical image segmentation. MedSAM2 (bowang-lab, April 2025) adds 3D and video. Neither is a Meddox chat model. | Open weights / research code. Not called by Meddox. |
| Merlin / CT-FM | CT / volumetric foundation models — not chat | Research / open weights where released |
| Prov-GigaPath | Pathology slide-scale foundation model — not chat | Research / Microsoft–Prov release channels |
| RadFM | Radiology foundation / multimodal imaging — not chat | Research / open weights where released |
Genomics / structure
| Name | Job | How you get it |
|---|---|---|
| AlphaFold 3 | Predicts how proteins interact with other molecules. Not an LLM. Not called by Meddox. | Non-commercial AlphaFold Server, academic weights, and a public structure database. The AlphaFold 3 code repository also describes a commercial Google Cloud path. |
| AlphaGenome Atlas | 8 September 2026: predicted effects of about 9 billion single-nucleotide variants, with an AVI score. Not an LLM. | Non-commercial portal. That post said commercial access via Google Cloud would come later. No Atlas API in Meddox. |
Therapeutics research
| Name | Job | How you get it |
|---|---|---|
| TxGNN-class models | Therapeutics / drug-repurposing research graphs | Research — not patient chat |
Leaderboards to watch
As of late 2026 — scores move. Outbound links only; we do not paste fake numbers.
- MedQA — classic medical QA (vals.ai / BenchLM / pricepertoken ecosystems)→
- MedXpertQA — medical QA with multimodal tracks→
- MedHELM — clinical scenarios, refreshed roughly quarterly→
- Medmarks — open medical evaluation suite→
- HealthBench — clinician-graded health conversations→
General frontier models often top these boards; open medical weights win on cost, privacy, and imaging-in-process.
What we plan to add
Some of this is still ahead. The wired ensemble is Gemini, Claude, and Grok.
- 1Keep MedGemma 1.5 off the chat picker. The imaging route does not call it.
- 2Specialty prompt packs are on the chat route. Diagnosis fields and ICD-10 ranges choose oncology, endocrinology, cardiology, or general internal medicine. The same three providers are called. This is a checklist, not a specialist model.
- 3Imaging path: stack viewer → key slices / volume → vision or MedGemma 1.5 — not 200 JPEGs into chat.
- 4Genomics: ingest a sequencing report; AlphaGenome Atlas as a research link only until a licensed commercial API exists.
AlphaGenome stays a research link. MedGemma is not a default route, and the published Google scores are not a Meddox result.