When the Model Imitates Its Teacher Instead of Learning the Task

Apertus 1.5 against Qwen3: both Apache 2.0, both fine-tunable. The most valuable finding from two days of measurement has nothing to do with the model — and everything with the fact that a model reliably learns whatever is in the data. Including the quirks of the model that generated it.

The question was simple: does a Swiss language model work as a base for a system of your own? Two days of measurement later the answer is yes — but the most valuable finding has nothing to do with the model. We tested Apertus 1.5 (ETH Zurich / EPFL) against Qwen3-8B. Same runtime, same quantisation, same task: turn prose into structured, schema-conformant output. Which is exactly what any system needs when it has to pull machine-readable information out of documents. First, a correction: both are equally open There is a persistent assumption that European models are "properly" open while Asian ones are merely "open weight" — that you can train the former but not the latter. That is not true. Both models are released under Apache 2.0, and both can be adapted via LoRA in the same way. We did it with Apertus; the identical path would have worked with Qwen3. The defensible difference lies in documentation, not licensing: Apertus publishes its training corpus — a public summary under Article 53(1)(d) of the EU AI Act. 17 trillion tokens, more than 1000 languages, knowledge cutoff April 2026. Apertus honours robots.txt opt-outs, retroactively. Anyone who objected to their content being used is excluded, even if the objection came later. The corpus has documented domain emphases, including legal texts and Swiss case law. Conversely, Apertus carries obligations Qwen3 does not: an acceptable use policy with an indemnification clause, a pass-through requirement on redistribution, and an output filter for personal data to be refreshed twice a year. In use it is therefore more restrictive, not more open. For organisations that publish themselves, or that must demonstrate what their system rests on, corpus traceability is the stronger argument. Just a different one than "trainable". What the measurement showed Without adaptation, Qwen3 produced valid, schema-conformant output in all four test cases. Apertus in one. The failures were consistently formal: wrong nesting despite balanced brackets, invented field names. The content was often not wrong — the model simply could not express the result cleanly. After fine-tuning, that entire error class disappeared: without fine-tuning with LoRA + schema constraint Apertus, valid output 1 of 4 24 of 24 Qwen3, valid output 4 of 4 — Add constrained decoding — a validator that masks, at sampling time, every token which cannot continue a valid prefix of the target schema — and the form is no longer likely but …

Author

Nexbid Engineering

Published

2026-08-15

Topics

Open Source, Research, Verification