開放模型
Apertus 1.5 Adds Native Image and Audio Understanding, With 8B and 70B Open Models Supporting 262K Tokens
The Swiss AI Initiative has released Apertus 1.5, using continued pretraining to expand its original text models into multimodal models that produce text output. The weights and training information are public, but the formal technical report, comprehensive benchmarks, and upstream inference framework support are still forthcoming.
The Swiss AI Initiative released Apertus 1.5 on July 24 in 8B and 70B variants. Rather than building a new architecture from scratch, the release retains Apertus 1.0’s decoder-only Transformer, xIELU activation, and AdEMAMix optimizer. The team added approximately 4 trillion and 2 trillion tokens of multimodal continued-pretraining data to the 8B and 70B models, respectively, enabling them to accept text, images, and audio while producing text output.
For deployment, the most immediate change is an increase in the default context length from the previous generation’s 65,536 tokens to 262,144 tokens. The release also adds a toggleable thinking mode, a tool-call parser, and more comprehensive instruction-following post-training. The model card provides vLLM launch instructions, with the 70B example using four-way tensor parallelism. However, users must currently run Swiss AI’s modified branches of vLLM or Transformers because the changes have not yet been officially merged upstream. A longer context window also does not mean every hardware configuration can use the full capacity out of the box. The project explicitly advises reducing `max-model-len` when memory is insufficient. Some multi-GPU configurations may also require fused operations to be disabled because of a conflict between fused all-reduce RMS and CUDA Graphs.
The technical significance of this release extends beyond its multimodal capabilities. Apertus 1.5 maintains the project’s commitment to Apache 2.0 licensing and to openly releasing model weights, data sources, training recipes, and intermediate artifacts, allowing governments, research institutions, and companies to audit and self-host the models. Engineering teams should watch three areas next: upstream vLLM and Transformers compatibility, measured throughput on image and audio tasks, and quality degradation at a 262K-token context length. The project says the complete technical report, benchmarks, training pipeline, and intermediate checkpoints will not be published for several more weeks. Until then, claims such as performance comparable to similarly sized models and improved reasoning cannot yet be independently reproduced.