- Deployment
- hosted
- Hq Country
- Switzerland
- Hq City
- Plan-les-Ouates, Geneva
- Ownership
- private-eu-owned
- Founded
- 2014
- Training Default
- never
- Residency Options
- Europe — "Lumo is built and based in Europe"; Proton AG is domiciled in Geneva and does not publish per-service data centre locations for Lumo
- Gov Access Exposure
- none-eu
- Eu Ai Act Role
- deployer
- DPA available
- yes
- Zero Retention Available
- Yes
- Retention Default
- "Our no-logs policy ensures we keep no logs of what you ask, or what Lumo replies." Saved chats are synced under zero-access encryption, so Proton holds ciphertext it cannot read.
- Retention Exceptions
- "Every conversation is zero-access encrypted, never logged, and never used to train future models." Proton's general privacy policy names third parties (Zendesk, Calendly, payment processors) for account and support functions, none of which sit in the inference path.
- Subprocessors
- Proton AG — "These run exclusively on servers Proton controls so your data is never stored on a third-party platform."
- Subprocessor Count
- 1
- Transfer Mechanism
- adequacy-decision
- Eu Procurement Ready
- conditional
- Eu Procurement Reason
- The strongest privacy architecture of any hosted assistant graded here — zero-access encryption on saved chats, a no-logs policy, open-source client code, models run only on Proton-controlled European servers — backed by a real DPA (updated January 2026) that confines transfers to Switzerland, the Union and adequacy-covered countries. It lands below the sovereign hosts for one reason: no certification. Proton publishes no ISO 27001, no SOC 2, nothing this pass could verify, which costs seven of the ten assurance points for a company whose entire proposition is security. It also declines to name the countries its Lumo servers sit in, saying only "Europe", so residency is recorded at the weaker `eu-default` rather than assumed generously. Worth noting for anyone reading sovereignty as provenance: the models behind Europe's most privacy-hardened assistant are Qwen 3.5 and GLM 5.2, both Chinese in origin. The rubric scores where the weights run and who can compel access, not who trained them — but the buyer should know.
- Training Claim Basis
- stated
- Training Confidence
- high
- Residency Claim Basis
- stated
- Residency Confidence
- medium
- Retention Claim Basis
- stated
- Retention Confidence
- high
- Subprocessors Claim Basis
- stated
- Subprocessors Confidence
- medium
- Residency Class
- eu-default
- Portability / Exit Path
- serves-open-models
- Subprocessor Jurisdiction
- eu-only
- Underlying model providers
- Alibaba (Qwen 3.5), Z.ai / Zhipu AI (GLM 5.2), open-weight image models (Image-Turbo, FireRed-Image-Edit-1.1)
