AI4SWITZERLAND
Research

ETH and EPFL Train Apertus 2.0 on Alps for a 2027 Code, Math and Agent Release

October 16, 2026 · 6 min read · Research

ETH and EPFL Train Apertus 2.0 on Alps for a 2027 Code, Math and Agent Release

Switzerland’s open Apertus language model is already being trained toward a 2.0 release on thousands of GPUs at the Alps supercomputer in Lugano, Blick reported after Apertus co-lead Imanol Schlag outlined plans at Zurich’s AI+X Summit. Schlag, of the ETH AI Center, co-leads the project with EPFL professor Martin Jaggi; they describe 2.0 as the largest model that can still be trained on Swiss-controlled infrastructure, with a first-quarter 2027 target and a further upgrade later in 2027.

Version 1.5 already handles images, audio and multi-step reasoning. The 2.0 brief prioritises coding, mathematics and agentic multi-step task completion while aiming to keep operating costs roughly flat. Schlag argued sovereign access matters after foreign models have been cut off for political reasons, citing Anthropic’s mid-year Claude restrictions under U.S. export controls as a cautionary case.

Everyday feedback is arriving via Proton’s Lumo assistant, where users can opt into sharing anonymised Apertus 1.5 chats with researchers, and via the Hack Apertus series that opened October 1 with an online round through October 16 before city events and a May 2027 St. Gallen final—all prototypes open-sourced.

What it means in practice

ETH and EPFL Train Apertus 2.0 on Alps for a 2027 Code, Math and Agent Release — contextual photo

Apertus 2.0 will be measured against risk politics such as MPs urging the Federal Council to prioritise extreme AI risks at the Geneva summit and engineering discipline from firms like Klarent, which raised CHF8 million for deterministic AI software testing. Consumer caution in Visa’s finding that nearly 70 percent of Swiss have used AI for shopping but few trust it to pay shows why agent features will need hard spend and action limits.

Cantonal IT buyers should budget evaluation time for Swiss-hosted open weights rather than assuming only U.S. APIs. Startups hacking on Apertus this month get a credible path into public-sector demos if they document safety refusals.

What to watch: Q1 2027 weights and eval cards, how many Hack Apertus prototypes reach production pilots, and whether federal procurement lists Apertus beside commercial models.

← Back to AI4Switzerland