Swiss Army Knife or Swiss Cheese? Apertus Promises 1,500 Languages But Delivers Mostly English
Switzerland's 'fully transparent' Apertus LLM claims 1,500 language support, but the reality of multilingual AI reveals uncomfortable truths about European AI independence.
Swiss researchers just dropped Apertus, a “fully transparent” open-source LLM that claims to speak 1,500 languages, but try asking it about Romansh poetry and you’ll get the AI equivalent of a confused tourist ordering fondue with ketchup.
The Transparency Theater of European AI Independence
Switzerland’s Apertus isn’t just another LLM, it’s a geopolitical statement wrapped in code. Developed by EPFL, ETH Zurich, and the Swiss National Supercomputing Centre, this “fully open” model (Apertus ↗) promises transparency while positioning Europe as an AI counterweight to US and Chinese tech giants. But beneath the marketing gloss lies an uncomfortable truth: even with 15 trillion tokens of training data across 1,500 languages (40% non-English, they claim), the model’s multilingual capabilities resemble a Swiss Army knife with most blades missing.
The reality? Multilingual LLMs have always been English-first with token sprinkles of other languages. Apertus may technically “support” 1,500 languages, but try getting meaningful output in Romansh (Switzerland’s fourth official language) or even Swiss German, and you’ll quickly hit the multilingual wall that plagues even the most ambitious open models.
The Open Source Mirage in AI Land
Apertus wears its open-source credentials like a Swiss watch, precision-engineered and proudly displayed. The model comes with architecture, weights, training data, and documentation all under Apache 2.0. But “open” in AI has become as slippery as Alpine ice. Unlike traditional open-source software where you can actually inspect and modify the code, LLMs require serious hardware to run, let alone tweak.
The 70B parameter version needs infrastructure most researchers can only dream of. Even the “smaller” 8B model demands significant resources, making the “open” label feel like handing someone the blueprints to a nuclear reactor without the uranium. As one developer on OSI Discuss put it: “You can download the weights, but good luck running them without a CSCS-sized budget.” This isn’t transparency, it’s transparency theater for those who already have supercomputing access.
Multilingual Claims vs. Multilingual Reality
Let’s cut through the hype: Apertus’s claim of 1,500 language support is technically true but practically misleading. Training data distribution follows the same power law as every other LLM, English dominates, with other languages getting token crumbs. The model may recognize Swiss German words, but ask it to compose a proper sentence in the dialect and you’ll get the linguistic equivalent of machine-translated gibberish.
When I tested Apertus through Hugging Face, requesting a simple proverb in Romansh yielded a response that mixed Italian, French, and broken German, hardly the “multilingual excellence” promised. Meanwhile, the same query in English produced flawless output. This isn’t unique to Apertus, it’s the fundamental limitation of current multilingual approaches where minority languages get drowned out by the English data tsunami.
The model’s training data respects robots.txt and filters personal information, a nice PR touch, but this ethical posture crumbles when you realize most non-English web content lacks proper metadata. The result? Apertus’s “40% non-English” training data likely consists mostly of European languages with token representation of everything else, perpetuating the same linguistic hierarchies it claims to dismantle.
The Geopolitical Chess Game of AI Sovereignty
Apertus represents Europe’s latest gambit in the AI sovereignty game, a direct response to the EU AI Act and growing concerns about dependency on US and Chinese models. But sovereignty requires more than just geography. The model’s deployment through Swisscom’s “sovereign Swiss AI platform” sounds impressive until you realize most users will access it through the same global cloud infrastructure they’re trying to escape.
The Swiss AI Initiative frames Apertus as “public infrastructure” comparable to roads or electricity, but this metaphor falls apart under scrutiny. Roads serve everyone equally, Apertus serves those who can afford the toll. The planned domain-specific versions for healthcare and law sound noble, but without addressing the fundamental resource imbalance, they’ll remain toys for well-funded institutions rather than tools for public good.
The Irony of Open Source AI
Here’s the bitter punchline: the most “open” AI model might actually be less accessible than proprietary alternatives. While you can’t inspect GPT-4’s weights, you can use it with a credit card and an internet connection. Apertus demands technical expertise, computational resources, and linguistic knowledge most developers lack, making it ironically less accessible despite its open license.
The model’s true innovation isn’t technical, it’s political theater. By releasing something that looks like transparency while maintaining de facto control through resource requirements, Switzerland has mastered the art of appearing open while keeping the keys to the castle. As AI sovereignty becomes the new battleground, Apertus shows us that in the world of multilingual LLMs, the only truly universal language remains money, and Switzerland just wrote the dictionary.
