Proton Integrates Apertus 1.5 into Lumo with Anonymous Feedback
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisis ton rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
The open Swiss model Apertus 1.5 arrives at Lumo with a system for anonymous feedback sent to researchers. Exchanges remain encrypted, and activation is at the user's discretion, while other services in Switzerland, such as Euria, also incorporate this model.
Euria also adopts Apertus, with distinct protections
Apertus is not exclusive to Lumo: Euria, Infomaniak's assistant, already offers it under the label "Swiss model in development," alongside two profiles based on Qwen 3.5. The guarantees are not identical from one service to another. Euria emphasizes a temporary operation, while Lumo applies zero-access encryption to stored conversations. According to Eamonn Maguire, AI director at Proton, having a European AI is not an end goal, and it must maintain its competitiveness to be effectively adopted.
In Lumo, opt-in and anonymous submission of negative feedback
Apertus 1.5 is not activated by default in Lumo: it must be manually selected, and users can revert to another model at any time. Without a rating assigned to a response, nothing is transmitted. A thumbs down anonymously sends the corresponding question and answer to Apertus researchers, and this feedback includes the text entered by the user. Proton has integrated Apertus 1.5 and announces that the model will leverage this feedback, with several tens of millions of potential Lumo users able to contribute on a voluntary basis.
An open model with public weights and Apache 2.0 license
Apertus claims to be fully open: its weights, training data, and methods are public, and it is distributed under the Apache 2.0 license. According to its creators, training was conducted exclusively using publicly accessible data, while respecting the removal requests from rights holders and excluding personal information. At its presentation in September 2025, it already had 70 billion parameters and had been trained on over 1,000 languages. This model results from a collaboration between ETH Zurich, EPFL, and CSCS.
Other models available and the issue of public feedback
Lumo already offered other models, including Qwen and GLM in its version 2.0. For Imanol Schlag from the ETH AI Center, the contribution of user feedback is the missing element for a fully open model funded by public means.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.