ETH and EPFL Train Apertus 2.0 on Alps for 2027 Release
ETH and EPFL co-leads told Zurich’s AI+X Summit around October 2, 2026 that Apertus 2.0 is already training on thousands of GPUs at the Alps supercomputer in Lugano, targeting a 2027 release—project pages flag the first quarter—with stronger coding, mathematics and agent capabilities while aiming to keep operating costs roughly flat versus today’s stack. Co-lead Imanol Schlag argued Switzerland needs open models that respect European rights because access to foreign systems is not guaranteed after export-control episodes such as Anthropic’s Claude restrictions, and the nationwide Hack Apertus series opened October 1 to push open-source prototypes through May 2027.
Filed under Research and dated October 4, 2026, this AI4Switzerland briefing treats the Apertus 2.0 build as Swiss sovereign-model news distinct from yesterday’s first public Iamo robotaxi pilot in Furttal. Version 1.5 already rides in Proton’s Lumo assistant with optional anonymised chat feedback, so the 2.0 cycle is about capability depth and real-world evaluation loops rather than a first launch.
Why it matters: open Swiss weights only matter if enterprises adopt them. A dated 2.0 roadmap plus hackathons can close the usage gap—if benchmarks and licensing stay explicit for regulated buyers.
What it means in practice
Swiss research, enterprise and counsel leads should inventory which workloads could move from foreign APIs onto Apertus 2.0 evaluation tracks; demand named owners for CSCS training milestones versus deployment assurance; assign an owner for how Lumo feedback sits beside Apertus 1.5’s Proton Lumo feedback loop; run time-boxed reviews of open-builder momentum already tracked beside the Open LLM Builders Summit at ETH; and prefer designs that keep humans on high-risk agent gates.
Caveats come first. Training start is not a Q1 2027 ship date; agent features need published evals; and parliamentary assistants still prefer commercial models today. AI4Switzerland therefore presents Apertus 2.0 as directional model-roadmap context until release candidates appear.
What to watch next: Hack Apertus online-round outputs through mid-October; published coding and math benchmarks for 2.0; and whether cantonal buyers pilot Apertus under the early-2027 AI-law consultation track. 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 1.5 feedback loops into a dated 2.0 training run and remains early. Organizations that benefit most will demand eval evidence, keep humans on agent gates, and refuse to confuse a summit teaser with finished sovereign stack.