Apertus 1.5 Joins Proton Lumo for Real-World Feedback Loop
The Apertus research team at ETH Zurich, EPFL and the Swiss National Supercomputing Centre partnered with Proton to integrate the open Apertus 1.5 language model into Proton’s privacy-focused Lumo assistant, ETH Zurich announced on September 17, 2026. The deal lets tens of millions of Lumo users optionally select Apertus and voluntarily share anonymised feedback—giving an academic open model a real-world improvement loop that proprietary chatbots have long enjoyed—while chats stay under Lumo’s zero-access encryption unless users opt in.
Filed under Research and dated September 28, 2026, this AI4Switzerland briefing treats the Proton integration as Swiss sovereign-model deployment news distinct from yesterday’s IBM Quantum System Two plan at CSCS Lugano. Co-leads Imanol Schlag and Martin Jaggi cast voluntary feedback as the missing ingredient for open models; Proton’s Eamonn Maguire argued Europe needs competitive domestic AI people will actually use, not merely research artefacts trained on Alps.
Why it matters: Swiss firms and cantons already experiment with Apertus on-prem and in the cloud. Consumer-scale feedback can harden quality—but only if consent logs, data-minimisation rules and human review of fine-tunes stay explicit.
What it means in practice
Swiss CIO and public-sector leads should inventory whether staff may select Apertus inside Lumo on managed devices; demand named owners for feedback-opt-in policies; assign an owner for comparing Lumo routes against on-prem Apertus; run time-boxed evaluations on Swiss-language tasks; and prefer designs that keep humans on model-update gates. Anchor the partnership to Ticino’s Apertus admin translation and Alps opening Apertus 1.5 on demand.
Caveats come first. Optional model selection is not default market share; open weights still trail frontier closed systems on some benchmarks; and partnership press is not a service-level guarantee for enterprises. AI4Switzerland therefore presents the Lumo link-up as directional research context until published usage metrics appear.
What to watch next: how many Lumo users enable Apertus; whether feedback datasets are released for research; and how Swisscom or Infomaniak offerings change. Readers can continue on the AI4Switzerland homepage, or browse the Newsroom for additional briefings.
Bottom line: treat this update as orientation, not instruction. Swiss open AI is moving from supercomputer releases into consumer assistants and remains early. Organizations that benefit most will treat consent as engineering, keep humans on fine-tune gates, and refuse to confuse a model selector with finished sovereignty.