A 20-billion-parameter language model fine-tuned on 5,000+ clinical scenarios. Runs locally inside your infrastructure. Patient data never leaves.
General-purpose AI models require sending patient data to external cloud servers. For regulated health systems, this creates compliance, sovereignty, and trust barriers.
Models trained on internet data produce confident-sounding but incorrect medical information. 52.1% overall diagnostic accuracy across general-purpose Large Language Models (LLMs).
Source: npj Digital Medicine 2025, systematic review of 83 studies
Per-token cloud pricing becomes unsustainable at national scale. Switching providers means rebuilding integrations from scratch.
Every request is classified before it reaches a model. Patient data stays local. Non-patient-data queries route to the best available model.
Translates between healthcare data standards — Health Level Seven version 2 (HL7v2) Admit/Discharge/Transfer (ADT) messages to Fast Healthcare Interoperability Resources Release 4 (FHIR R4) Patient and Encounter resources. Handles the 80% of routine transformations that would otherwise require manual mapping.
Routes referrals by clinical priority, agent capability, and capacity. Integrates with the Global Healthcare Agent Registry and Routing Architecture (GHARRA) for cross-organisational agent discovery.
Differential diagnosis suggestion from symptoms, observations, and patient history. Part of the ensemble diagnostics system — Bayesian, heuristic, LLM, and expert rules voting together to reduce single-model risk.
Maps between Systematized Nomenclature of Medicine — Clinical Terms (SNOMED CT), International Classification of Diseases (ICD-10/11), Logical Observation Identifiers Names and Codes (LOINC), RxNorm, and the dictionary of medicines and devices (dm+d). Enables laboratory observations to flow across protocol boundaries.
Drug interaction checking, allergy alerts, contraindication detection, dosage verification, and serotonin syndrome risk assessment using Hunter criteria. Safety gates that never compromise for speed.
The model router classifies every request. Patient data stays with Bevan. Non-patient-data queries cascade to external models only when Bevan cannot handle them with sufficient confidence.
| Concern | Cloud-Only AI | Bevan (Local-First) |
|---|---|---|
| Patient data exposure | Data leaves premises | Stays local — always |
| Latency | Network round-trip | Sub-millisecond |
| Cost at scale | Per-token billing escalates | Fixed infrastructure cost |
| Regulatory compliance | Complex data processing agreements | Data residency by default |
| Availability | Internet dependency | Runs offline |
| Customisation | Prompt engineering only | Fine-tuned on your domain |
| Vendor lock-in | Single provider dependency | Multi-model with local fallback |
“The NHS will last as long as there are folk left with the faith to fight for it.”
— Aneurin Bevan, 1948
Bevan LLM is named after Aneurin Bevan, the Welsh politician who founded the United Kingdom’s National Health Service in 1948 — the principle that healthcare should be available to everyone, regardless of wealth. That same principle drives this model. Healthcare AI reasoning should be available to every health system, not just those that can afford proprietary cloud AI. A local-first model that runs on your infrastructure, fine-tuned on your clinical domain, governed by your policies. That is what Bevan represents.
See the platform in action
See how Bevan LLM delivers domain-specific reasoning inside your infrastructure — with full audit trails and no patient data leaving your premises.