Team Apertool · Hack Apertus 2026

Four tools.
One open model.

Apertool is a one-person team building four tools on Apertus, Switzerland's fully open language model. Like the blades of a pocket knife, they fold out of the same handle: one tests the model, one measures it, two put it to work.

Apertool logo: a pocket knife unfolding four tools, one per project
Each tool in the knife is one project. Tap a tool to jump to it.

About Apertus and Hack Apertus

Apertus

Apertus is a large language model from EPFL, ETH Zurich and CSCS, released fully in the open: weights and training data. It was trained across 1,811 languages, and its name is Latin for open. Every Apertool project runs on Apertus v1.5.

Hack Apertus

An open-source hackathon series by the aiLights Association to evaluate, align and build on Apertus. The online stage runs from 1 to 16 October 2026 with four challenges in two tracks: readiness and adoption. Apertool enters all four.

The tools

Four blades, one handle.

Each card is one blade of the knife. Track 1 gets Apertus ready, Track 2 puts it to use.

Track 1A · Red-Teaming Apertus

CheeseDrill

Problem
The Swiss AI Charter sets out how Apertus should behave. CheeseDrill tests where those promises give way on everyday Swiss questions.
Solution
Charter-based red teaming: paired probes in German, French and Italian test Apertus against its own charter, and one command reproduces every run.

Results go to the Apertus team first and stay confidential until 1 December 2026, as the track's disclosure rules require.

Track 1B · Swiss Voices

Inkraft

Problem
Swiss social insurance law changes on fixed dates, and a model cannot tell which version it learned, so last year's rule comes back as a confident answer.
Solution
A trilingual benchmark (DE · FR · IT) whose gold answers rest on verbatim fedlex quotes, telling correct, stale and invented answers apart.

Dataset v1.0.0: 326 question cores, 948 rows, drafted by language models and checked automatically; a human graded a blind 22 % sample of the judgements (κ 0.91). Apertus 70B answers 51 % of rows correctly, but only 7 % of the rows on amended rules with a stored earlier answer, where it gives the earlier law in 30 %.

Track 2A · Academia Challenge by OST

GeaNaVid

Problem
Voters hear a claim, but the answer is buried somewhere in the official voting booklet.
Solution
Apertus checks each claim against the booklet and marks it Gea, Na or Vid (Romansh for yes, no, blank); a Gea or Na quotes the booklet line behind it.

OST benchmark v1.1, test split: macro-F1 0.959 with Apertus 8B and 0.965 with 70B, against 0.720 and 0.889 when the model reads the whole booklet; 0.874 and 0.932 on unseen booklets (synthetic hold-out).

Track 2B · Own Project

Inbrev

Problem
Every morning a Swiss GP practice faces one inbox of lab reports, refill requests, insurer letters and faxes in three languages, all under professional secrecy.
Solution
Inbrev sorts the inbox by handling path and deadline with Apertus on a box inside the practice, and every field it extracts quotes its source.

180 synthetic messages, labels checked by one person on a sample of 60 (κ 0.98 route, 1.00 clinical content): route macro-F1 0.94 (8B) and 0.96 (70B), keyword rules 0.76; all 16 messages with clinical content leave automation (8B with its second check; keyword rules 7).

Video

The knife, unfolded.

A short tour of all four tools, from the first probe to the sorted inbox.

How we work

Open model, open work.

Open

Code, data and reports are published as each track's rules allow. Red-teaming results follow responsible disclosure and stay private until 1 December 2026.

Sovereign

Apertus does the language work in every tool. We build for Swiss infrastructure or hardware you own, not for a foreign cloud API.

Reproducible

Each repo is built around one command in Docker: make run. CheeseDrill replays its recorded runs without an API key; GeaNaVid and Inbrev run against an Apertus endpoint; Inkraft recomputes its results from the stored answers and judgements without an API key.

Measured, not made up

No number goes on a slide before a run or a cited source backs it. Until then, it says TBD.

Contact

Questions, ideas or feedback? Write to Team Apertool and we will answer by e-mail.

Your name, e-mail address and message are forwarded by e-mail to Team Apertool, solely to answer you. The form does not store them and does not pass them on.