The Apertus project announces Apertus 1.5, an updated version of its 8B and 70B models. It adds image and audio understanding, an optional reasoning mode, a context window four times longer, and improved performance in instruction following and tool use. The weights for the 8B and 70B models were published on Hugging Face on July 24, 2026, under the Apache 2.0 license.
Apertus 1.5 extends the training of the first generation. According to the project, 4 trillion tokens of text and multimodal data were added to the 8B model, and 2 trillion to the 70B model, for a total of 19T for the former and 17T for the latter. A token is a unit of text, segmented by the model for processing. As with previous versions, the project states that the models are fully open: weights, data, value charter, and training details are published.
Images, Audio, and Expanded Context
Images and audio enter the model as discrete tokens, placed in the same sequence as text, not via a separate visual encoder. Images are segmented using the Emu3.5 tokenizer, with a 512×512 pixel image becoming 1,024 tokens, and audio with WavTokenizer, at a rate of 40 tokens per second. The project specifies that the models understand images and audio but only generate text. Speech support remains experimental.
The context window, meaning the amount of information the model can process at once, reaches 262,144 tokens, compared to four times less for Apertus 1.0. The reasoning mode, which can be activated, allows the model to think before responding, which according to the project improves its results on logic problems.
Post-Training and Evaluations
Post-training takes place in five steps: supervised fine-tuning (SFT) on 3.7 million text samples, supplemented with image and speech data, a second fine-tuning on long context, reinforcement learning with verifiable rewards (RLVR) on approximately 350,000 queries, direct preference optimization (DPO), then alignment with the Apertus charter.
According to the project, the 8B model surpasses comparable-sized open models in instruction following and multilingual tasks, while still lagging behind the best cutting-edge open models. Multimodal performance is said to approach that of Gemma 4 12B. On the Last Translation Benchmark, the 8B model reportedly matches GPT-OSS-20B in machine translation and outperforms Claude Haiku 4.5 and Google Translate, based on published results.
Several providers already offer Apertus 1.5. The Proton Lumo and Infomaniak Euria applications have integrated it, and Swisscom, Infomaniak, and Phoeniqs have announced API access. The project also indicates it has surpassed five million downloads according to its Hugging Face statistics. At the AI+X summit in Zurich, the team presented the roadmap for Apertus 2.0, which is to be based on a sparse MoE-type architecture, have around 600 billion total parameters, and prioritize agents, mathematics, and code.
