Apertus 1.5 makes evaluation conditions part of the release
Multimodal inputs arrive alongside intermediate checkpoints and structured records of how evaluations were run.
A benchmark score is easier to interpret when its generation settings and execution conditions travel with it. Apertus 1.5 makes that reporting visible through evaluation records using the Every Eval Ever schema, alongside its expanded multimodal capabilities.
Hanna Yukhymenko and colleagues at Swiss AI published the release notes on October 15, 2026. The weights had already been released on July 24. These are distinct dates. The Apache 2.0 models accept images and audio as well as text, while generating text only. Spoken-language processing remains experimental.
The team also provides intermediate checkpoints and documents the training stages. My preferred research starting point would be a controlled comparison across those checkpoints, keeping prompts and generation settings fixed. That would help distinguish changes introduced during training from differences caused by evaluation setup.
Structured records make comparisons inspectable; they do not establish performance on every downstream workload. The open question is which reported improvements survive a change in language, document type or recording conditions, particularly for the experimental audio capability.