Apertus 1.5: open models with 262,144-token context — four times longer than before
The Swiss AI Initiative has updated its open language models with image and audio understanding, an optional thinking mode, and a context window four times as long as before. The release is a concrete example of what "fully open" can mean in practice — but all capability claims so far come from the project's own release notes.
The Swiss AI Initiative published the release notes for Apertus 1.5 on Hugging Face on October 15, 2026. The notes document an update whose model weights were made available earlier, on July 24, 2026, under the Apache 2.0 license. Two dates are thus involved: the weight release in July and the subsequent documentation in October, without the notes describing exactly how the two relate.
The release is interesting precisely because it, according to the project, documents a model far beyond the weights: the notes describe the Apertus models as "fully open" — with open weights, open data, open values, and complete training details. This is the project's own description, not independent verification, but it illustrates how much the scope of openness can vary between releases that are all called "open."
What's new
According to the release notes, Apertus 1.5 extends the first-generation 8B and 70B models with several capabilities:
- Understanding of images and audio as input
- An optional thinking mode
- A context window four times as long as in Apertus 1.0
- Improved instruction following
- Improved tool use
All of these points are the project's own claims from the release notes. No independent measurements confirming them are available in the accessible source material, and without such data the model cannot be compared against specific closed or other open models.
The technology: continued pretraining, not a rebuild
Apertus 1.5 was not trained from scratch. According to the notes, it is continued pretraining of the first-generation models: 4 trillion tokens of text and multimodal training data were added to the 8B model, which has now seen a total of 19 trillion tokens. The 70B model received an additional 2 trillion tokens, for a total of 17 trillion.
This is a technically meaningful distinction. Continued pretraining builds on an already-trained model rather than repeating the entire training run from the start. For people already using Apertus 1.0, this means the starting point is the existing models — not new architectures or entirely new model families.
How the multimodality works
Image and audio understanding is introduced through what the project calls early fusion. Instead of attaching a separate vision encoder to the model, images and audio are converted into discrete tokens that are inserted into the same sequence as text tokens. According to the notes, the project uses the Emu3.5 image tokenizer and WavTokenizer for audio.
The practical consequence is that the model architecture remains unchanged — only the vocabulary is extended with tokens for the new modalities. It is this mechanism that, according to the project, makes it possible to extend an existing text model to multimodal input through continued pretraining, without changing the model structure itself.
The context window: from 8k to 262,144 tokens
The context window has been quadrupled from Apertus 1.0, to 262,144 tokens. According to the notes, this was not achieved in one step: the window grows incrementally from 8k tokens in a dedicated long-context training phase.
For users, a 262,144-token window means the model can in principle work with considerably longer documents and conversation histories than before. How well it exploits the full window in practice cannot be determined from the available source material.
Limits and caveats
Several limitations are worth noting:
- The models understand images and audio but generate only text. This is not multimodal output.
- Spoken-language processing is described as experimental.
- The claims about capabilities and improvements come from the project's own notes and are not confirmed by any third party in the available material.
That last point is perhaps the most important for prospective users: it remains to be seen whether the claimed improvements in instruction following, tool use, and multimodal understanding hold up in practice.
Background: open versus closed
The release lands in the middle of an ongoing debate about open versus closed models, where the question of what "open" actually means is central. In an industry where "open" often means only open weights — without training data or training details — Apertus 1.5 is a concrete data point for what more comprehensive openness can involve: open weights, open data, and complete training details, under an Apache 2.0 license.
That description rests on the project's own self-characterization and should not be read as an independent assessment of the quality of the openness or of the model's competitiveness. But as a case, the release shows how a research initiative chooses to document a model release — and thus provides a concrete example in a debate that is otherwise often conducted in abstract terms.
Source: Apertus 1.5 release notes (09d5ec19).