A model’s final score cannot explain which training decision produced its behavior. That is a problem for teams trying to adapt a model without losing capabilities they already rely on. Access to intermediate stages makes a different kind of investigation possible: comparing what changed before and after a specific intervention.
The Swiss AI Initiative’s October 15 Apertus 1.5 notes describe 8-billion- and 70-billion-parameter models whose weights were released on July 24. The documentation covers image and audio understanding, optional thinking and a longer context window. It also links data records, training recipes and intermediate checkpoints. The October publication is additional technical documentation, rather than the first availability of those weights. Primary source.
A visible sequence of changes
According to the authors, images and audio become discrete tokens alongside text, while generated output remains text. Continued pretraining precedes several post-training stages, including supervised learning, reinforcement learning and preference alignment. Spoken-language processing remains experimental. The release notes acknowledge that the models still trail frontier open-weight systems in some comparisons.
For an adopting team, the value of these records is the possibility of isolating a regression. If a task becomes worse after an alignment stage, an intermediate checkpoint gives investigators a concrete comparison. That does not establish the cause by itself: decoding settings, evaluation data and serving software also have to remain controlled.
Openness as an experimental resource
The useful first experiment is narrower than replacing a production assistant. Choose a task with known answers, preserve its inputs and evaluate multiple stages under identical conditions. Inspect failure types as well as aggregate scores. A model that follows instructions more consistently could still lose a specialized capability that matters to that application.
Longer context and additional input modalities expand the testing burden. An image may be legible to a person but encoded poorly; a long document may fit without its relevant detail being used. Those are separate capabilities to measure.
Public training artifacts could make adaptation and independent analysis more accountable if outside teams can reproduce the reported behavior. The next evidence is such reproduction on clearly specified tasks, including regressions and resource costs. Documentation creates the opportunity for scrutiny; it is not itself proof that a deployment is ready.
References
Questions, news tips or corrections? Contact [email protected].
