Open-weight models let hospitals keep the data and self-host the intelligence
Self-hostable medical models are here: Google's MedGemma 1.5 (a small, open, fine-tunable clinical model, 2026-01-13) plus enterprise clinical assistants like OpenEvidence (~$12B, ~20M consultations/month by early 2026). Combined with K3/Inkling as fine-tune bases, a hospital or biotech can tune a frontier-class model on proprietary clinical data inside its own walls — PHI never leaves the building. The defensible asset is the private data loop, not the base model.
Why it matters (spine — health). Human health, measured honestly. Spec an on-prem clinical fine-tune reference (privacy-preserving, audit-logged) for HKG; sell the verification/ops layer, not the model.
What would prove it wrong (falsifier). On-prem clinical fine-tunes underperform hosted frontier APIs enough that health systems keep sending data out.
Live-status check + triangulation. Self-hostable medical models are here: Google's MedGemma 1.5 (a small, open, fine-tunable clinical model, 2026-01-13) plus enterprise clinical assistants like OpenEvidence (~$12B, ~20M consultations/month by early 2026). Combined with K3/Inkling as fine-tune bases, a hospital or biotech can tune a frontier-class model on proprietary clinical data inside its own walls — PHI never leaves the building. The defensible asset is the private data loop, not the base model. Source: Company · Google (MedGemma 1.5) / OpenEvidence (2026-01-13); see source link. Every post-cutoff figure was tool-fetched against a live source during the Kimi K3 sweep (2026-07-20); see sweeps/2026-07-20-VERIFIED_DIGEST.md for the full provenance and any corrections of record.
Spec an on-prem clinical fine-tune reference (privacy-preserving, audit-logged) for HKG; sell the verification/ops layer, not the model.
domains: health · clinical-ai · measurement