Swiss AI releases Apertus 1.5 with vision, audio and longer context

The open-weight 8B and 70B models add multimodality, but still trail frontier rivals

Swiss AI has released Apertus 1.5, an update to its open-weight 8B and 70B language models. It adds image and audio understanding, an optional thinking mode, and a context window four times longer than its predecessor's. The updated weights went live on Hugging Face on July 24, 2026 under an Apache 2.0 licence. A technical report details the full training pipeline, data sources and evaluation methodology.

The release is built on continued pretraining from Apertus 1.0. It adds 4 trillion tokens of text and multimodal data to the 8B model and 2 trillion to the 70B, bringing their totals to 19 trillion and 17 trillion tokens respectively. Images and audio enter the model as discrete tokens alongside text, using the Emu3.5 tokenizer for images and WavTokenizer for audio, rather than a separate vision encoder. Swiss AI says this choice keeps the architecture unchanged and training at text-only speed. The context window grows to 262,144 tokens through a staged pretraining process that moves from 8,000 tokens up to the full length. More than half of all training tokens remain text-only, to preserve the original model's language ability.

Post-training runs through five stages: supervised fine-tuning, long-context fine-tuning, reinforcement learning with verifiable rewards on about 350,000 prompts, preference optimisation, and a final charter-alignment stage called SDPO. Swiss AI reports that Apertus 1.5 8B now beats similarly sized open-weight models on instruction-following and multilingual benchmarks. Its machine translation matches GPT-OSS-20B, the company says, while outperforming Claude Haiku 4.5 and Google Translate on the Last Translation Benchmark. Swiss AI says multimodal performance sits closer to Gemma 4 12B and remains competitive in spatial and document understanding among open-weight peers. But it acknowledges the models still trail frontier open-weight systems overall.

Swiss AI publishes its evaluation results in the Every Eval Ever schema, developed with the EvalEval Coalition. It records who ran each test, how the model was accessed, and generation settings. The company notes the UK AI Security Institute also uses this infrastructure for its public evaluations. Swiss AI says the schema can represent 56 of 60 reporting elements called for under Article 55 of the EU AI Act and the GPAI Code of Practice. Apps Proton Lumo and Infomaniak Euria have integrated Apertus 1.5. API access is available through Swisscom, Infomaniak, Phoeniqs and other providers. At the AI+X Summit in Zurich, the team outlined a roadmap for Apertus 2.0, a sparse mixture-of-experts model with roughly 600 billion total parameters that will prioritise agents, mathematics and coding.

Written from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.