Apertus 1.5 continues training from version 1.0 with trillions of extra tokens of text and multimodal data. Images and audio are turned into tokens in the same stream as text, so the architecture stays the same; the models still only produce text, and speech input is experimental. The weights are under the Apache 2.0 licence, and the team says all new training data was publicly available and permissively licensed, with website opt-outs respected.
Why it matters
Most capable models are open in their weights only. Apertus is one of the few where the data and training process are open as well, which matters to researchers, regulators and organisations that need to know what a model learned from. Version 1.5 narrows the feature gap with less transparent models.
Original source
Apertus 1.5 Release Notes
Hugging Face
huggingface.co
Summary by Second Order AI. Full reporting is at the original source.