Apertus 1.5 Release Notes

Hugging Face’s Apertus project has shipped Apertus 1.5, a sizable update to its 8B and 70B models. The new weights landed on July 24, 2026 under Apache 2.0, and the pitch is clear: keep the models fully open while making them do more than plain text. That now includes images and audio, plus an optional thinking mode and a context window that stretches to 262,144 tokens.

This is not a fresh model from scratch. Apertus 1.5 is continued pretraining on top of Apertus 1.0, with 4 trillion extra tokens for the 8B model and 2 trillion for the 70B model. The project says the total training mix now stands at 19T and 17T tokens, respectively. More than half of that is still text-only, which is meant to keep the original language ability intact.

The multimodal part is built in an unusually direct way. Images and audio are turned into tokens and dropped into the same sequence as text, rather than routed through a separate vision stack. The models understand image and audio inputs, but they still output text only. Hugging Face says the 512×512 image is tokenized into 1,024 tokens, while audio uses 40 tokens per second. The architecture itself stays unchanged.

The training recipe is split into stages: unpaired text, images and audio; paired image-caption and audio-transcript data; a cooldown phase with higher-quality math and code; and a long-context stage that grows the window from 8k to 262k. Post-training then layers on SFT, long-context SFT, RL, DPO, and charter alignment. The project also says it used only publicly available, permissively licensed data, with robots.txt opt-outs respected retroactively.

On benchmarks, Apertus 1.5 is still behind frontier open-weight models, but the gains are real. The 8B model outperforms similarly sized open-weight models on instruction-following and multilingual tasks, and thinking mode helps on reasoning-heavy tests like AIME. The multimodal results are said to be close to Gemma 4 12B and other current open-weight models, with especially strong showing on spatial understanding and document/chart tasks. The project also points to Swiss Legal and translation results, and says its evaluation reports are published in a machine-readable format built with the EvalEval Coalition.