Swiss AI Initiative releases Apertus 1.5, adding multimodal input and a 4x longer context window

The Swiss AI Initiative — a joint effort of ETH Zurich, EPFL, and the Swiss National Supercomputing Centre (CSCS) — released Apertus 1.5, an update to its fully open language model family that adds image and audio understanding and quadruples the model's context window.

What's new

The model card on Hugging Face describes the release plainly: "Apertus 1.5 is a family of 8B and 70B parameter language models designed to advance the state of multilingual, multimodal, fully open, and transparent AI." Multimodality is new to the line: "As a first in the Apertus family, the Apertus 1.5 models support multimodal inputs. The model takes images, audio, and text as input and generates text."

Context length also jumped significantly: "Apertus 1.5 by default supports a context length up to 262,144 tokens, a four-fold increase from our initial Apertus 1.0 release." The 8B and 70B models were built via continued pretraining of Apertus 1.0 — adding a multimodal mix of roughly 4 trillion tokens for the 8B model and 2 trillion tokens for the 70B model — on the same decoder-only transformer architecture as the original release. Both models are released under the Apache 2.0 license and are available now on Hugging Face under the swiss-ai organization, alongside optional thinking-mode reasoning before generating a response.

Context

Apertus 1.0 launched in September 2025 as one of the most fully documented open-weight model efforts to date — publishing not just weights but training data, code, and development process, an unusually high transparency bar even among open-weight releases. It was trained on the Alps supercomputer at CSCS, a Swiss government-funded system, making Apertus one of the few frontier-adjacent open models built entirely on sovereign European infrastructure rather than a hyperscaler's cloud.

Why it matters

Apertus 1.5 keeps Switzerland's model in step with the broader shift toward multimodal-by-default open releases from labs like Meta and Mistral, while preserving the transparency commitments that set it apart — full training-data and pipeline disclosure is still rare even among open-weight competitors. For European institutions and enterprises weighing data sovereignty, a fully open, Apache-licensed model trained on domestic compute is a meaningfully different option than open-weight models originating from US or Chinese labs, even when the underlying capabilities are comparable.