Apertus 1.5 folds images and audio into the text vocabulary: same architecture, 8B and 70B weights open under Apache 2.0
Synopsis
Switzerland's Apertus project released 1.5 versions of its 8B and 70B models, continuing pretraining from 1.0 to add image and audio understanding, an optional thinking mode, and a context window of 262,144 tokens. Multimodality arrives by early fusion: images and audio become discrete tokens placed in the same sequence as text, with no separate vision encoder, so the architecture is unchanged and training runs at text-only speed. The models read images and audio but generate only text, speech handling remains experimental, and the project says the release still trails frontier open-weight models.
Interpretation
Apertus 1.5 lets the 8B and 70B text models take in images and audio directly, and switch into a thinking mode that reasons before answering. Apertus 1.0 handled text only and had a much shorter context; 1.5 grows the window to 262,144 tokens, four times that of 1.0. The project reports that the 8B version outperforms similarly sized open-weight models on instruction following and multilingual tasks, and that multimodal performance is closer to Gemma 4 12B and other current open-weight models, competitive in spatial understanding among open-weight models and in document and chart understanding among open-source models.
Multimodality does not come from an attached vision encoder; images and audio become the same kind of discrete token as words. The common route trains a separate vision encoder and connects it to the language model; here only the vocabulary is extended and the architecture stays unchanged. Images are tokenized with Emu3.5, where a 512×512 image becomes 1,024 tokens, and audio with WavTokenizer at 40 tokens per second, so training still runs at text-only speed.
Continued pretraining proceeds in four stages, from unpaired data to paired data, then a higher-quality cooldown and a stepwise long-context extension. The starting point is the Apertus 1.0 weights, with 4 trillion tokens added to the 8B model and 2 trillion to the 70B model, reaching 19T and 17T in total. More than half of the training tokens remain text-only to preserve the language abilities of 1.0, and new data follows the same bar of being publicly available, permissively licensed, and respecting robots.txt opt-outs retroactively, with every source listed in the data transparency record.
Perspective
This report draws on the Apertus 1.5 release notes and covers the 8B and 70B sizes, the four-stage continued pretraining and five-stage post-training, and evaluations of text, multimodal and thinking modes. It suits researchers and engineers interested in open-weight training pipelines and how multimodality is wired in. The next step stated in the release notes is the Apertus 2.0 roadmap: a sparse/MoE architecture with roughly 600B total parameters, an extended training schedule, and agents, mathematics and coding as priorities.
The release notes say the new version still trails frontier open-weight models, that speech handling remains experimental, and that the models generate text only. Multimodal and thinking-mode evaluations are organized by the project itself and published through EvalEval's evaluation-card format with run settings; comparisons with Gemma 4 12B and others are the project's own statements, and specific scores should be checked against the evaluation records.