Apertus 1.5 models add image and audio understanding and a context window of 262 144 tokens
Apertus 1.5 models in 8B and 70B variants add image and audio understanding; output remains text-only. The context window expands fourfold to 262 144 tokens. The authors are releasing weights, data preparation and training procedures, and evaluation suites under permissive licenses.
The authors have released Apertus 1.5 models in 8B and 70B variants. The family, previously focused on text, now supports image and audio understanding but still produces only text output. The context window expands fourfold to 262 144 tokens. According to the authors, the architecture is unchanged and a tokenizer handles multimodal inputs.
Development builds on Apertus 1.0 models through continued pretraining on an additional 4 trillion tokens for the 8B variant and 2 trillion for the 70B variant. Reasoning mode is enabled by a flag in the chat template; it is not a separate model. According to the authors, it improves mathematical reasoning scores for Apertus 1.5 8B by up to 20 points.
Model weights, data preparation and training procedures, and evaluation suites have been released under permissive licenses, according to the authors. The authors report the strongest multimodal results in document understanding and speech recognition in long recordings or in languages with less available data. They also acknowledge that specialized image and audio models remain better at perception tasks and transcription in English and other languages with more available data.
Why it matters
The release expands the open model family to handle image and audio inputs while retaining text responses. The longer context window supports larger inputs, and the released weights and procedures give developers resources for their own deployments and evaluations.
Two audiences, two different impacts
What this means
For individuals
There is another option when choosing a tool for document understanding or speech recognition. However, according to the results reported by the authors, Apertus 1.5 models are not the best choice for every task, especially English transcription.
For a business
Development teams gain access to weights, training procedures and evaluation suites for a multimodal system. This allows them to assess deployment for their own business tasks; the authors cite permissive licenses, but the source does not specify the particular licensing terms.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.