The editorial argues Apertus ships not just weights but training data, curation pipeline, training code, and evaluation harness — every byte needed to reproduce the model. By contrast, Llama restricts commercial use above 700M MAUs and hides its corpus, Mistral gated its strongest models, and DeepSeek released weights but not data. Apertus is framed as the first well-funded model that meets the OSAID 1.0 definition without compromise.
The editorial positions Apertus as Europe's answer to 'open-ish' US models by emphasizing that it was trained on the Alps supercomputer at CSCS in Lugano, by a Swiss public research consortium (EPFL, ETH Zürich, CSCS), with a license that doesn't claw back rights based on user count. Sovereignty here means provenance throughout the stack, not just hosting weights locally.
By submitting Apertus under the framing 'Open Foundation Model for Sovereign AI' and driving it to 398 points, the submitter highlights the sovereignty angle as the core appeal — a model whose entire supply chain sits within Swiss/EU jurisdiction rather than depending on US-controlled labs.
The editorial notes that GPAI providers crossing the 10^25 FLOP systemic-risk threshold now face transparency obligations that US labs are 'quietly stonewalling.' Apertus inverts this posture by disclosing everything rather than the minimum, suggesting that a public-research-consortium model is structurally better suited to the AI Act's regime than a venture-funded lab that needs to monetize.
Apertus — Latin for 'open' — landed on Hacker News at 398 points, a Swiss-built foundation model from EPFL, ETH Zürich, and the Swiss National Supercomputing Centre (CSCS). Unlike Llama 3, Mistral, or Qwen, Apertus ships not just the weights but the training data, the curation pipeline, the training code, and the evaluation harness — every byte needed to reproduce the model from scratch.
The model was trained on the Alps supercomputer, a GH200-based system at CSCS in Lugano, and is being positioned as Europe's answer to the 'open-ish' models that have dominated 2024–2025. The framing is explicit: this is sovereign AI, meaning the entire supply chain — data provenance, compute, weights, governance — sits inside Swiss and EU jurisdictions, with a license that doesn't claw back rights based on user count or downstream use.
The release timing matters. The EU AI Act's general-purpose AI obligations have been phasing in through 2025–2026, and providers of GPAI models with 'systemic risk' (the 10^25 FLOP threshold) now face transparency requirements that most US labs are quietly stonewalling. Apertus inverts that posture: rather than disclose the minimum, it discloses everything.
The word 'open' has been doing increasingly load-bearing work in AI marketing. Meta calls Llama open while restricting commercial use above 700M MAUs and never releasing the training corpus. Mistral started open and pivoted to a freemium gate around its strongest models. DeepSeek shipped weights but not data. The Open Source Initiative's [OSAID 1.0 definition](https://opensource.org/ai/open-source-ai-definition) — finalized last October — requires training data, code, and weights, and by that bar, almost nothing on Hugging Face actually qualifies.
Apertus is the first well-funded model that meets the OSAID bar without asterisks, and it's being shipped by a public research consortium rather than a startup that will inevitably need to monetize. That's the structural difference. EPFL and ETH don't need to convert open-source goodwill into Series B momentum. CSCS isn't burning runway. The incentive to renege on openness later — the trajectory that took Stability, Mistral, and arguably OpenAI itself from genuinely open to nominally open — doesn't exist here.
The benchmark question is the obvious one, and the honest answer is: Apertus is not going to beat Claude 4.5 or GPT-5 on MMLU-Pro. That's not the point. The point is that for a large class of European developers — those at banks under DORA, hospitals under the EHDS, public-sector contractors, anyone whose legal team has spent the last 18 months reading the AI Act — the choice isn't between Apertus and GPT-5. It's between Apertus and *not deploying generative AI at all because procurement can't sign off on a US-hosted black box*.
The HN thread surfaced the predictable critiques: parameter count modest by 2026 standards, multilingual coverage skewed toward European languages, inference still expensive without a quantized release. All fair. But the comparison matters because a 70B fully-open model that you can audit, retrain, and host on European hardware is categorically different from a 400B closed model accessed via an API in us-east-1, even if the closed model scores ten points higher on a benchmark. Compliance is not a benchmark you can ace at inference time — it's a property of the entire pipeline, and Apertus is the first model where that pipeline is actually inspectable.
There's also the geopolitical undertow. The 2025 export control fights, the on-again-off-again chip restrictions, the question of whether Anthropic or OpenAI would honor European data residency commitments under a sufficiently aggressive US administration — all of this has made 'sovereign AI' shift from a buzzword used by national champions chasing subsidies into an actual procurement requirement. France's Mistral was supposed to be that answer; it has spent the last year drifting toward the same closed-API business model as its US competitors. Apertus, by virtue of its non-commercial origins, can't drift the same way.
If you're building agents, RAG systems, or fine-tuned classifiers and you've been on Llama 3.1 or 3.3 because the license was tolerable: Apertus is worth a serious eval week. Pull the weights, run your domain benchmarks, and specifically test the long-tail of non-English European languages where Llama is weakest — that's where Apertus's training data curation is likely to show its edge. If your eval delta against Llama is less than 5 points on tasks you care about, switching is a free legal upgrade.
If you're at a regulated EU company that's been waiting for an answer to 'what model can we self-host that our DPO will sign off on,' the answer just arrived. Run it on your own GH200s, or rent CSCS time, or use one of the European inference providers that will inevitably stand up Apertus endpoints in the next 60 days. The AI Act Article 53 transparency obligations are dramatically easier to meet when the provider has already published everything.
If you're an open-source maintainer building tools — fine-tuning libraries, evaluation harnesses, agent frameworks — Apertus is the most ethically defensible base model to demo against. Llama demos always come with the asterisk that someone's lawyer hates. Apertus demos don't.
The interesting question isn't whether Apertus catches the frontier — it won't, and that's fine. The question is whether the *next* fully-open model from this consortium, or its inevitable Chinese, Japanese, or Indian counterparts, can be trained on enough compute to close the gap. The 2026 fight is no longer about whether open models can compete with closed ones on quality — DeepSeek settled that — but whether truly open models can compete with nominally-open ones on adoption. Apertus is the first datapoint in that fight, and the fact that it came from a Swiss research consortium rather than a venture-funded lab tells you something about which institutions still have the freedom to ship something genuinely open.
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