Helvetra moves to Apertus 1.5: what changes, and what we're watching

On 24 July 2026, the team behind Apertus (ETH Zurich, EPFL and the Swiss National Supercomputing Centre) released Apertus 1.5. It is not a new model built from scratch but a continued training of the one Helvetra already used: the same Swiss, fully open foundation, carried further. For the 70B model that powers our translations, that meant two trillion additional tokens of training on top of version 1.0.

On 4 August 2026, Infomaniak added Apertus 1.5 to the AI infrastructure we run on, and we pointed Helvetra's translation calls at the new version. For us it was a small change: one model name swapped for another, and a restart. For you it was invisible: no downtime, nothing to do, and from that moment the translations simply come from 1.5.

The update is about capability and reliability more than raw size. The parts that matter:

• A much longer memory. The context window grows fourfold, from 64k to 262,144 tokens, so the model can hold far more text at once.
• Tighter instruction-following. The updated post-training makes the model stick more precisely to what it is told, with more predictable output.
• It can now see. Version 1.5 accepts images alongside text. Audio input exists too, though the team still calls it experimental.
• Better tool use and an optional reasoning mode, aimed at developers building agents on top of the model.

A full technical report and benchmark results are due in the coming weeks, so the precise numbers are not public yet. We are not going to quote figures that do not exist.

Not every headline feature matters for a translation tool, so here is the honest split. The one most likely to help you is the tighter instruction-following. Helvetra asks the model to do exactly one thing, translate, and nothing else. A model that obeys that more reliably means cleaner results and fewer stray, chatty, or off-target answers. The longer context also makes it realistic to translate much longer passages in a single pass than we comfortably handle today.

Some of the new abilities we will probably leave switched off, and we would rather say so. The reasoning "thinking mode" is built for hard problem-solving, not translation, and turning it on would mostly make you wait longer for the same sentence. Image input is genuinely interesting, translating the text inside a screenshot or a photo is an obvious future feature, but Helvetra is text-only today and we are not going to pretend otherwise.

This is the one most of you actually care about, and it is where we are most careful. Swiss German is where Apertus has been weakest, and we have said so plainly before. Version 1.5 does not publish a Swiss German number yet, the language-by-language benchmarks are part of a technical report still to come, so we did the next best thing: on the day we switched, we ran Apertus 1.5 through Helvetra's own translation prompts and watched how it behaved. What follows is a first look, not a benchmark.

The encouraging part: it produces Swiss German dialect rather than refusing. Ask for Bernese and "Wie bisch hüt?" comes out as you would hope. Formality holds in the other languages too, du and Sie, tu and vous. And, importantly for a translation tool, it still resists prompt injection: give it a sentence with a hidden instruction and it translates that instruction rather than obeying it. That last point matters more than it sounds, because it is what keeps Helvetra doing only its one job.

The rough edges are the ones you would expect. It sometimes leaves a set greeting like "Grüezi mitenand" untranslated, and telling Swiss German apart from standard German automatically is still unreliable, a clearly dialectal sentence can be read as plain German. So 1.5 is a step, not a finish line. The real measurement is still you: if you translate into Swiss German, use the thumbs up or thumbs down and tell us. You are better placed to judge your own dialect than any score we could quote.

A note on how we got here, because it says something about why we built Helvetra the way we did. Every layer of Helvetra runs on Infomaniak, the Swiss host whose independence we wrote about last month, and Apertus, the Swiss open model we covered after Uphill Conf. When 1.5 came out, we did not just wait and hope. We wrote to Infomaniak directly to ask when it would reach their AI API. The answer came back fast, and the wait turned out to be short.

That is the quiet advantage of a small, sovereign stack. The people who build the model publish it in the open, the Swiss host we chose puts it on tap, and a one-person tool can pass it on to you within a day of it going live. No layer in that chain answers to anyone outside the country. When the next version lands, we will do the same thing again.

• The announcement: Apertus 1.5 (apertus-ai.org)
• The model we run: Apertus v1.5 70B (Hugging Face)
• The write-up: Building the next generation of open AI infrastructure (CSCS)
• Background on the model: Apertus at Uphill Conf 2026
• Why we host on Infomaniak: Infomaniak becomes a foundation