ETH Zurich's Apertus 1.5 Adds Image and Audio to Its Open Model Family
Switzerland's fully open Apertus family grows eyes and ears, with a 262K context window, optional thinking mode, and Apache 2.0 weights for everything.
- ETH Zurich, EPFL and CSCS released Apertus 1.5, fully open 8B and 70B multimodal models under Apache 2.0.
- Continued pretraining added 4T tokens (8B) and 2T tokens (70B) for image and audio understanding.
- Context window quadrupled to 262,144 tokens with no architectural changes from Apertus 1.0.
- Thinking mode is a chat-template flag, boosting 8B math reasoning by up to 20 points.
- Weights, data, recipes and evaluation suites are all public on Hugging Face.
- Strongest on document understanding and low-resource speech; specialist models still lead on English ASR and perception.
Apertus 1.5 adds image and audio input to Switzerland’s open model family
ETH Zurich, EPFL and the Swiss National Supercomputing Centre have released Apertus 1.5, extending the text-only Apertus family with image and audio input. The 8B and 70B models generate text, support up to 262,144 tokens of context and retain the decoder-only transformer architecture from Apertus 1.0.
Two sizes, one open stack
| Specification | 8B model | 70B model |
|---|---|---|
| Inputs | Text, images and audio | Text, images and audio |
| Output | Text | Text |
| Additional pretraining | 4 trillion tokens | 2 trillion tokens |
| Maximum context | 262,144 tokens | 262,144 tokens |
| License | Apache 2.0 | Apache 2.0 |
Both checkpoints result from continued pretraining of Apertus 1.0 and retain its transformer blocks, xIELU activation and AdEMAMix optimizer. The release also includes training data, recipes, evaluation suites and intermediate checkpoints, giving researchers the artifacts needed to inspect and adapt the system.
Public funding, inspectable artifacts
Apertus publishes material frequently absent from model releases, including the training mixture and intermediate states. The project’s definition of openness covers weights, data, code and training recipes, which allows outside groups to trace model behavior to specific design and data decisions.
That approach has already supported derivative work. Estonia’s EstLLM project at the University of Tartu and EPFL’s Meditron medical group both publish models built on Apertus weights.
The data policy carries over from Apertus 1.0: training uses openly available sources, excludes restrictively licensed material and applies crawler opt-outs retroactively when preparing the corpus. These records can support provenance reviews and regulatory documentation, though adopting the model does not by itself establish legal or EU AI Act compliance.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
