Apertus 1.5 models add thinking mode; speech processing is experimental
The Apertus 1.5 models have a switchable thinking mode; speech processing remains experimental. The 8B variant underwent further training on 4 trillion tokens, and the 70B variant on 2 trillion. The totals reached 19 and 17 trillion tokens.
The release details for the Apertus 1.5 models describe continued pretraining: the 8B variant adds 4 trillion tokens for a total of 19 trillion, while the 70B variant adds 2 trillion for a total of 17 trillion. Thinking mode can be enabled through a chat template setting; it is not a separate model. According to the authors, it improves performance on tasks requiring reasoning. Spoken language processing remains experimental.
The released variants build on the Apertus 1.0 models and accept text, images and audio, but generate only text. The context window has quadrupled to 262 144 tokens. Image and audio inputs are converted into tokens in the same sequence as text, while the model architecture remains unchanged.
The authors are releasing weights, data preparation and training procedures, and evaluation datasets under permissive licenses. According to their evaluations, instruction following, coding and the use of external tools and APIs have improved. They also acknowledge that specialized image and audio models remain better at some perception and speech transcription tasks.
Why it matters
Developers can combine the analysis of text, documents, diagrams, photographs and audio within a single family of open models. The switchable thinking mode allows them to try both response modes without changing models; however, experimental speech processing requires validation for the specific use case.
What was added since the original report
Verified updates
-
The 8B variant adds 4 trillion training tokens for a total of 19 trillion.; The 70B variant adds 2 trillion training tokens for a total of 17 trillion.; Spoken language processing is experimental.; The models can be switched into thinking mode.; The models were trained for better integration of external tools and APIs.
- The 8B variant adds 4 trillion training tokens for a total of 19 trillion.
- The 70B variant adds 2 trillion training tokens for a total of 17 trillion.
- Spoken language processing is experimental.
- The models can be switched into thinking mode.
- The models were trained for better integration of external tools and APIs.
Two audiences, two different impacts
What this means
For individuals
When working with recordings, you can obtain a text response, but the experimental status of speech processing is a reason to check its accuracy against the original recording.
For a business
Teams preparing their own deployments gain access to model weights as well as procedures for data preparation, training and evaluation. This gives them the resources to review data provenance and conduct reproducible model evaluations.
DevelopmentCheck the original
Event sources
independently confirmed · 2 publishers, 1 independent. We count feeds from the same owner only once.