Apertus 1.5: Switzerland's fully open model reads images and audio

ETH Zurich, EPFL and CSCS released Apertus 1.5 on 24 July: fully open 8B and 70B models with image and audio input, a 262K context and Apache 2.0. Audio understanding is experimental.

  • New release
  • 2 min read
  • Published

ETH Zurich, EPFL and the Swiss National Supercomputing Centre (CSCS) released Apertus 1.5 on 24 July 2026. It is a family of 8B and 70B models that now accepts images and audio as well as text, with a 262,144-token context window, an optional thinking mode and an Apache 2.0 license.

What is new

Apertus 1.5 continues the pre-training of the first-generation Apertus 1.0 models on a further 4 trillion tokens for the 8B model and 2 trillion for the 70B, for totals of 19T and 17T tokens. The architecture is unchanged: a decoder-only transformer with the xIELU activation, trained with the AdEMAMix optimizer. Images and audio enter as discrete tokens in the same sequence as text, through an Emu3.5-based image tokenizer and WavTokenizer, instead of through a separate vision encoder, and the models generate text only. The context window is four times that of Apertus 1.0, and thinking is a flag of the chat template, not a separate model.

In the technical report the 8B model leads fully open models of its class on knowledge, instruction following and cultural benchmarks, while specialized vision and audio models stay ahead on perception and on transcription of English and other high-resource languages. The project describes audio understanding as experimental. These are the developers’ own evaluations.

Running it

The BF16 weights are about 18.4 GB for the 8B model and 144.6 GB for the 70B. At release the model card said support was not yet in released Transformers or vLLM versions and pointed to Swiss AI’s forks and a Docker image. The Hugging Face repositories are gated, so you must accept the access terms before you download. The card says no output filter is provided, so add your own safeguards, and it recommends refreshing the weights about every six months to pick up data-deletion requests. Tool calling does not work in thinking mode, and video input is not supported, so extract frames and pass them as images.

License

Apertus is open in the fullest sense. The weights, intermediate checkpoints, data preparation pipelines, training code and evaluation harnesses are all published under Apache 2.0. The team says it trains only on openly available data, honors crawler opt-outs retroactively and excludes restrictively licensed material. Base and post-training stage checkpoints (SFT, DPO and RL) are published as separate repositories.

Why it matters

Most open-weight models publish the weights and little else. Apertus and OLMo 3.1 are the two models in our catalog tagged fully open, and Apertus is the only one that reads images and audio. That matters to public bodies, universities and companies that must document where a model’s training data came from, for example for EU AI Act work, and to researchers who want to reproduce training with intermediate checkpoints at hand. It does not lead on raw capability: Qwen3.8-27B and Gemma 4 are the models to compare it with, and the report itself says specialized models stay ahead on perception and on transcription of high-resource languages.

Open weights vs open source explains the difference. Apertus is also on the Europe list and the multimodal models list.

Sources

Models in this story

Model Released Parameters Context License Smallest variant needs
Apertus 1.5 Swiss AI Initiative 8B – 70B 256K Apache 2.0 (Permissive) 5.52 GB Laptop or phone
Olmo 3.1 Allen Institute for AI (Ai2) 7.3B – 32.2B 64K Apache 2.0 (Permissive) 5.04 GB Laptop or phone
Qwen3.8-27B Alibaba Qwen 27.8B 256K Apache 2.0 (Permissive) 19 GB Gaming GPU
Gemma 4 Google DeepMind 5.1B – 30.7B 256K Apache 2.0 (Permissive) 3.52 GB Laptop or phone