Apertus 1.5 Release Notes
huggingface ·
The Apertus team has released version 1.5 of their open-weight language models, featuring both 8B and 70B parameter variants under an Apache 2.0 license. This update introduces native multimodal capabilities, allowing the models to process images and audio alongside text through a unified tokenizer architecture. The context window has been expanded fourfold to 262,144 tokens, and an optional thinking mode has been added to enhance reasoning performance on complex tasks. Training involved continuing pre-training with trillions of additional tokens across unpaired, paired, and long-context stages, followed by rigorous post-training phases including supervised fine-tuning and reinforcement learning. Evaluation results indicate significant improvements in instruction following, tool use, and multilingual translation, with the 8B model outperforming similar-sized competitors on specific benchmarks. The project emphasizes full transparency by publishing all training data, recipes, and intermediate checkpoints on Hugging Face. Additionally, Apertus collaborates with the EvalEval Coalition to provide reproducible evaluation cards that comply with emerging AI governance standards like the EU AI Act. Looking ahead, the team has announced a roadmap for Apertus 2.0, which will focus on sparse architectures and agent-based capabilities.
huggingface