Apertus 1.5: early-fusion images and audio, a 262k context and every training stage open
Victoria Shinder3 min readThe Swiss AI team behind Apertus has published release notes for Apertus 1.5 on Hugging Face, alongside a technical report (arXiv:2610.15542, dated 11 October 2026). The updated 8B and 70B weights were released on 24 July 2026 under the Apache 2.0 licence; the new post summarises how they were trained and evaluated.
What 1.5 adds. Apertus 1.5 is a continued pretraining of the first-generation models. The 8B model received 4 trillion additional tokens of text and multimodal data and the 70B model 2 trillion, for totals of 19T and 17T. The release adds image and audio understanding (audio is described as experimental), an optional thinking mode, a context window of 262,144 tokens — four times that of Apertus 1.0 — and better instruction-following and tool use. The models understand images and audio but generate only text.
Early fusion instead of a vision encoder. Images and audio enter the model as discrete tokens in the same sequence as text. Images are tokenised with the Emu3.5 tokenizer, so a 512×512 image becomes 1,024 tokens; audio uses WavTokenizer at 40 tokens per second. Because only the vocabulary is extended, the team says the architecture is unchanged and training runs at text-only speed. Continued pretraining ran in four stages — unpaired data, paired image–caption and audio–transcript data, a cooldown with more maths and code, and a long-context stage growing the window from 8k to 262k tokens. More than half of the training tokens are text-only. The data rules stay as before: publicly available, permissively licensed, with robots.txt opt-outs respected retroactively, and every source listed in the apertus-data-transparency repository.
Post-training in five stages. SFT on 3.7 million text samples plus image and speech data; long-context SFT at the full window; reinforcement learning with verifiable rewards on about 350,000 prompts; DPO on judge-scored preference pairs; and, for the 8B model, alignment with the Apertus Charter through SDPO. The SFT text mix, the preference pool and the preference pairs are published, and intermediate checkpoints — base, SFT, RLVR and DPO — are in the project's Hugging Face collection.
How it compares. The team writes that the models still trail frontier open-weight models, but that the 8B model outperforms similarly sized open-weight models on instruction-following and multilingual tasks, and that its multimodal performance is closer to Gemma 4 12B. External evaluators report Swiss legal tasks on par with leading open-weight models and, on a translation benchmark, the 8B model matching GPT-OSS-20B.
Evaluation records. Results are published as Evaluation Cards in the Every Eval Ever (EEE) schema of the EvalEval Coalition, which records who ran an evaluation, how the model was accessed and the generation settings. The post cites EvalEval's finding that the schema can represent 56 of 60 evaluation-reporting elements in Article 55 of the EU AI Act and the GPAI Code of Practice.
What is next. The team reports five million downloads and a roadmap for Apertus 2.0: a sparse mixture-of-experts architecture of about 600 billion total parameters, with agents, mathematics and coding as priorities.
Source: Hugging Face Blog, "Apertus 1.5 Release Notes" — https://huggingface.co/blog/swiss-ai/apertus-v15-release-notes