Imanol Schlag

@ImanolSchlag

Co-invented DeltaNet. Leading the development of Apertus: a fully open and trustworthy LLM developed in CH. Lecturer of "Large Scale AI Engineering" at ETHZ.

Switzerland

ischlag.github.io

Replies Reposts Media

Imanol Schlag
@ImanolSchlag
Sep 17
Open models rarely get the most important step of development: real users & real feedback. Apertus 1.5 is now in Lumo, @ProtonPrivacy 's privacy-first assistant. Chats stay encrypted; users can choose to donate anonymised feedback to our team. Let's go!
Apertus partners with Proton’s Lumo AI assistant
From ethz.ch
1
Imanol Schlag
@ImanolSchlag
Jul 27
Our friends at publicai.co have built a dialogue on what people in Switzerland want from AI. It takes about ten minutes, it's anonymous, and the results go into an open report. Your input will help inform the direction we take with Apertus. dialogue.publicai.ch
Imanol Schlag
@ImanolSchlag
Jul 24
Hey 👋 Apertus 1.5 is now available to the public on publicai.co! 😍 Chat with it here: chat.publicai.co Try out the image capabilities! I'd be curious what people think before sharing detailed benchmark results :)
Public AI Inference Utility
From publicai.co
Imanol Schlag
@ImanolSchlag
Jul 24
Today we are releasing Apertus 1.5 8B and 70B, a continued pretraining of our Apertus 1.0 models, which have passed 3.2M downloads in the community. 🧵
1
Imanol Schlag
@ImanolSchlag
Apr 28
🚨From our Apertus RL work, @AlexisLimozin found two bugs in DeepSpeed and OpenRLHF that suppress SFT baselines. Affects several recent mixed-policy RL-for-reasoning papers. With fixes, SFT-then-RL still wins. w/@valentina__py, Eduard Ďurech, @thoefler arxiv.org/abs/2604.23747
Alexis Limozin
@AlexisLimozin
Apr 28
We found two bugs in DeepSpeed and OpenRLHF that deflated SFT baselines in several recent mixed-policy RL papers. Once fixed, plain SFT-then-RL beats or matches every mixed-policy method we tested.
1