Apertus 1.5 Release Notes

Apertus 1.5 extends the 8B and 70B Apertus 1.0 models with image and audio understanding, an optional thinking mode, a four-times longer context window, improved instruction-following, and better tool use. The updated model weights were released to Hugging Face on July 24, 2026 under an Apache 2.0 license. The release includes a Technical Report detailing enhancements and training processes, emphasizing full openness with open weights, data, values, and training details.

Multimodality is introduced through early-fusion tokenization, where images and audio are treated as discrete tokens alongside text tokens. Images are tokenized with the Emu3.5 tokenizer, converting a 512×512 image into 1,024 tokens, while audio uses WavTokenizer at 40 tokens per second. The architecture remains unchanged, allowing training to proceed at text-only speed. Over half of the training tokens remain text-only to preserve language abilities from Apertus 1.0.

Post-training for Apertus 1.5 follows five stages: supervised fine-tuning, long-context SFT, reinforcement learning, direct preference optimization, and charter alignment. The release expands evaluations to include multimodal and thinking-mode benchmarks, using frameworks like MLLM-eval-suite and lm-evaluation-harness. Apertus 1.5 8B outperforms similarly sized open-weight models in instruction-following and multilingual tasks, while multimodal performance is competitive with models like Gemma 4 12B.

The Apertus 1.5 results are published in the Every Eval Ever (EEE) schema, which records evaluation settings, metrics, and metadata for transparency. The release includes intermediate checkpoints, training data mixes, and model cards on Hugging Face, enabling researchers to reproduce and build upon the work. Collaborations with providers like Swisscom and Infomaniak have led to deployments, with API access announced by multiple companies.