Switzerland's Apertus 1.5 Gains Vision, Hearing and a 262K-Token Context Window

The Swiss Apertus project has released Apertus 1.5, an update to its open models at 8B and 70B parameters. The models now accept images and audio, can switch into a thinking mode, work with a context of 262 144 tokens (four times that of Apertus 1.0), and are better at calling external tools. The weights were released on Hugging Face on 24 July 2026 under Apache 2.0, and a detailed technical report was published in October. This is continued training of the first version: the 8B model received another 4 trillion tokens of text and multimodal data (19 trillion in total), and the 70B received 2 trillion (17 trillion in total).

The release's main distinguishing feature is how the team added multimodality. It is also fully open: the weights, data, training recipes and values charter are all publicly available.

Multimodality without a separate encoder

Instead of the usual setup with a vision encoder, the developers chose early fusion. Images and audio are converted into discrete tokens and run in the same sequence as text. Emu3.5 tokenizes images: a 512×512 image yields 1 024 tokens. WavTokenizer processes audio at 40 tokens per second. The architecture stays the same and only the vocabulary expands, so training runs at the speed of a text-only model. There is a limitation: the models understand images and audio but generate only text. Speech handling is still considered experimental.

Pretraining ran in four stages: unpaired data (text plus standalone images and audio), paired data (image captions and transcripts), a cooldown focused on math and code, and context extension from 8k to 262k tokens. More than half of the tokens remained purely text, so the models would not lose the language skills of Apertus 1.0. The text corpus was refreshed with FineWeb2-HQ+, FinePDFs and Nemotron-CC. The conditions are unchanged: only public data under permissive licenses, and robots.txt opt-outs are honored retroactively.

Five stages of post-training

After pretraining, the models go through a chain of SFT, long-context SFT, RL, DPO and charter alignment (SDPO).

• SFT on 3.7 million text examples plus image and speech data. Reasoning traces are added only to prompts the model cannot solve without them. Tool-use data is limited to trajectories verified by execution.
• SFT on the full 262k window restores the ability to work with long context, which short-context training weakens.
• RLVR with verifiable rewards on roughly 350 thousand prompts covering math, code, instruction following, multilingual tasks and vision, with the thinking mode both on and off.
• DPO on judge-labeled pairs improves chat quality, tool use and safety.
• SDPO aligns the model with the Apertus Charter through judge feedback and self-distillation. This stage was applied only to the 8B.

In ablations, the team noticed something interesting: multimodal data in SFT improves results across all modalities. And the gain from multimodal reasoning traces survives the subsequent text-only RL: math improves more than when starting from a purely text checkpoint.

What the benchmarks show

The authors state candidly that the models still trail leading open-weight solutions. Even so, the progress is noticeable.

• Apertus 1.5 8B outperforms open models of similar size on instruction-following and multilingual tasks.
• The thinking mode raises results on hard benchmarks such as AIME.
• In multimodality, the model is closer to Gemma 4 12B and other current open-weight solutions. Among fully open models, it is competitive in spatial understanding and in working with documents and charts.
• On Swiss legal tasks, the model performs on par with leading open-weight models.
• On the Last Translation Benchmark, the 8B version matches GPT-OSS-20B and beats Claude Haiku 4.5 and Google Translate.

The team puts particular emphasis on transparent evaluation. Results are published in the Every Eval Ever format together with the EvalEval Coalition: the record captures not only the score but also who ran the test, how the model was accessed, and which generation settings were used. According to EvalEval, the scheme covers 56 of the 60 reporting elements required by Article 55 of the EU AI Act and the GPAI Code of Practice. The tests run on a public fork of lm-evaluation-harness, so they can be rerun.

Where to get it and what comes next

The release includes base models, SFT, RL, DPO and final checkpoints for the 8B and 70B (the final 70B corresponds to the DPO stage). Hugging Face hosts intermediate checkpoints, the text SFT mixture, the preference pool and the training pairs. The code builds on Megatron-LM and verl. Proton Lumo and Infomaniak Euria have already integrated the models. Swisscom, Infomaniak, Phoeniqs and other providers have announced APIs.

For European companies with strict requirements on data provenance and compliance, this is one of the few options where the entire path from corpus to checkpoint can be verified. Researchers have access to every stage for study. Downloads on Hugging Face already exceed five million. The team outlined its next step at the AI+X Summit in Zurich: Apertus 2.0, with a sparse MoE architecture of roughly 600B parameters, a longer training schedule, and a priority on agents, math and code.