The editorial argues Apertus is the first model at this scale where 'what was this trained on?' can be answered with a citation list instead of a shrug. Unlike Llama 3, Mistral, and DeepSeek which ship only weights, Apertus publishes the full training corpus recipe and training code, making it qualitatively different from other 'open' models.
The editorial contends that the August 2026 EU AI Act enforcement deadline forces general-purpose model providers into disclosure requirements that closed-weight US models can't satisfy without revealing trade secrets. European deployers using OpenAI or Anthropic for high-risk applications are building compliance programs around a black box, and Apertus is the first credible answer to that gap.
The editorial highlights that Apertus trained on 1,800+ languages with explicit emphasis on Swiss German, Romansh, and other low-resource European languages systematically underweighted by US-trained models. It also retroactively honored robots.txt and ai.txt opt-outs by removing content from sites that signaled 'no AI training' after the crawl — described as a first in the industry.
By submitting Apertus to Hacker News, T-A signaled that this Swiss open foundation model deserves developer attention. The 330-point score and 112 comments validate that the HN community broadly agreed this release is significant enough to elevate above competing stories.
EPFL, ETH Zurich, and the Swiss National Supercomputing Centre (CSCS) released Apertus, an open foundation model trained on the Alps supercomputer in Lugano. The release ships in two sizes — 8B and 70B parameters — and the headline isn't the parameter count. It's the license stack underneath it.
Apertus publishes three things that almost no other 'open' model publishes together: the weights, the full training corpus (or rather, the documented, filterable, opt-out-respecting recipe to reconstruct it), and the training code. Llama 3 ships weights under a custom license. Mistral ships weights. DeepSeek ships weights. Apertus is the first model at this scale where you can answer 'what was this trained on?' with a citation list instead of a shrug.
The model was trained on roughly 15 trillion tokens across 1,800+ languages, with explicit emphasis on Swiss German, Romansh, and other low-resource European languages that have been systematically underweighted in US-trained foundation models. The team built the corpus by respecting `robots.txt` and the new `ai.txt` opt-out conventions retroactively — websites that signaled 'no AI training' after the crawl had their content removed from the training data. That's a first.
The 'sovereign AI' framing is easy to dismiss as marketing. It isn't. Under the EU AI Act, general-purpose model providers face disclosure requirements that effectively no closed-weight US model can satisfy without revealing trade secrets — and the August 2026 enforcement deadline is ten weeks out. European deployers using OpenAI or Anthropic for high-risk applications are now in the position of building compliance programs around a black box. Apertus is the first credible answer.
Compare the provenance stories. With Llama 3, you know it was trained on '15 trillion tokens of publicly available data' — that's the entire disclosure. With GPT-4, you don't even get that. With Apertus, you get a structured manifest of source domains, a documented filter pipeline, and a reproducibility guarantee: the same code and the same corpus on the same hardware produces the same weights. For a Swiss bank or a German hospital evaluating a model for patient-facing use, this isn't a nice-to-have. It's the difference between a 6-week procurement and an 18-month one.
The benchmark story is more interesting than the press release suggests. On standard English benchmarks — MMLU, GSM8K, HumanEval — Apertus 70B tracks roughly with Llama 3.1 70B and well behind Llama 3.3 70B. On low-resource European languages, it's not close: Apertus beats every US-trained open model on Romansh, Swiss German, Basque, and Catalan evaluations, often by 15-30 points. The lesson buried in those numbers: when you stop treating English as 92% of the training mix, multilingual performance stops being an afterthought. That's a real architectural insight, not a marketing one.
The HN thread (currently at 330 points) is split in the predictable ways. One camp argues this is what 'open' should have meant all along, and that Meta's Llama license has been allowed to define the term down. Another camp points out that 70B is too small to compete on the agentic tasks where frontier capability actually matters in 2026, and that 'sovereign' AI built on PyTorch (US-controlled) and CUDA (US-controlled) running on Nvidia silicon (US export-controlled) is sovereign in name only. Both are right. The model isn't a GPT-4o replacement; it's a regulatory replacement.
If you're running inference for European customers, three things change this quarter. First, your DPIA gets shorter. Apertus has a documented training corpus, so the 'unable to determine training data provenance' line in your Article 35 assessment becomes a citation instead. Second, your model card gets honest. You can publish what your downstream system was trained on without a lawyer review cycle. Third, your fine-tuning options expand — because the training code is published, you can resume pretraining from a checkpoint rather than just LoRA-ing the final weights, which matters if you're trying to teach the model a domain (medical, legal, financial) where vocabulary drift hurts.
The deployment story is also unusually clean for an academic release. Apertus ships GGUF quantizations day-one, vLLM-compatible weights, and a reference Triton serving config. The 8B model runs on a single H100 at full precision and on a 24GB consumer GPU at Q4 — which means the gap between 'this exists on Hugging Face' and 'this is in production' is measured in hours, not weeks. That's not always true for academic releases.
The gotcha: there's no commercial support contract behind it. CSCS is a public supercomputing center, not a vendor. If your incident response runbook for an LLM regression involves 'page our model vendor,' Apertus doesn't fit that workflow yet. Expect a Swiss commercial spinout within 12 months — the same pattern as Mistral emerging from Meta's open-weights momentum in 2023 — but right now, you're on your own for production support.
The long bet here isn't that Apertus wins benchmarks. It's that the AI Act's transparency requirements, combined with the slow-motion data-residency squeeze on US cloud providers in Europe, makes 'fully open and provenance-documented' a structural moat rather than a hippie commitment. If that bet is right, Apertus is the first model of a category that will dominate European regulated deployments by 2028 — and every closed-weight provider will eventually have to ship a 'transparency tier' to compete. If it's wrong, this is a beautifully engineered academic project that 200 European researchers will love and nobody else will deploy. The next 18 months — and the AI Act's first wave of enforcement actions — will tell us which.
I like the idea, and it has become more pressing that everyone outside the US think about tech sovereignty because the US has become an unsafe place to keep your data, but the impression I get from Apertus is that it moves at the speed of a committee. I have no expectation they'll deliver a com
By far the most impactful product of the Apretus project are the people. To quote a memorable line from Dominique Paul (https://www.thisiscrispin.com/):> What most people miss IMO is that this is not a team who is doing this for the fourth time like virtually any other LLM provider
For a model that claims to focus on many languages, it's quite unreliable when it comes to simple questions like "how to say X in language Y" or "how to conjugate verb X in language Y". It keeps hallucinating words that do not exist, and when corrected, it only hallucinates
Looks like their instruct models are Llama3.1 fine tune from last year. Is there any progress on new models?My last hope for soverign AI is from Chinese open models
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
Other fully open LLMs include Allen AI's OLMo 3.1 and MBZUAI's K2 Think V2, both of which have released their full training pipelines and datasets.Nvidia Nemotron is also an open training source model, though a portion of its dataset remains proprietary.Quoting lambda's comment:> N