Most open-weight models let you download the finished weights but not see what went into them. Apertus, from the Swiss AI Initiative, is one of the few families where the training data sources, recipes and intermediate checkpoints are public too, and version 1.5 makes it considerably more useful. The 8B and 70B models now read images and audio, can think before answering, and handle a context of 262,144 tokens, four times longer than Apertus 1.0. Everything remains under Apache 2.0.
The method is simple in a useful way. Rather than bolting on a separate vision encoder, the team turned images and sound into tokens in the same stream as text, a technique called early fusion. A 512 by 512 image becomes 1,024 tokens and audio runs at 40 tokens a second. Only the vocabulary changed, so the architecture is the same and training runs as fast as text-only. Continued pretraining added 4 trillion tokens to the 8B model and 2 trillion to the 70B, with more than half still text so that language skills hold. Post-training ran in five stages, including reinforcement learning on about 350,000 prompts with checkable answers.
On results, the authors are candid. They say Apertus 1.5 still trails the leading open-weight models. The 8B beats models of similar size on following instructions and on multilingual tasks, its multimodal scores come close to Gemma 4 12B, and on one translation benchmark it matches GPT-OSS-20B and beats Claude Haiku 4.5 and Google Translate. Those are the authors' own evaluations, published in charts and an evaluation card rather than as numbers in the text.
The transparency is the point. New data is publicly available and permissively licensed, robots.txt opt-outs are honoured retroactively, every source is listed in a public repository, and the team publishes the training summary required under the EU AI Act. For public bodies, universities and companies that need to answer questions about what a model was trained on, that is worth more than a few benchmark points. The models are available through Swisscom, Infomaniak and other providers, and in apps such as Proton Lumo, and Apertus has passed 5 million downloads in total. Gemma 4 and GPT-OSS-20B, which we review, are stronger general choices; Apertus is the one you can audit.
The limits are clear. Spoken-language input is described as experimental. The models read images and audio but only write text. Performance trails frontier models, which is the usual cost of training only on openly licensed data. The blog does not say how many languages are covered beyond calling the models globally multilingual. The next step is more ambitious: Apertus 2.0 is planned as a mixture-of-experts model of about 600B parameters, which will test whether full openness can scale.