Apertus 1.5 Adds Multimodal AI and 262,144-Token Context to Open Models
Apertus 1.5 adds image and audio understanding, thinking mode and a 262,144-token context to open 8B and 70B models, with APIs and full training data.
Summary
Apertus published Apertus 1.5 details on October 15, 2026, after releasing Apache 2.0 8B and 70B weights on Hugging Face on July 24 and an arXiv report on October 11. Continued pretraining added 4 trillion tokens to 8B and 2 trillion to 70B, bringing totals to 19T and 17T. Both gain image and experimental audio understanding, optional thinking, stronger instruction following and tool use, and a fourfold 262,144-token context window. Emu3.5 turns a 512 by 512 image into 1,024 tokens, while WavTokenizer uses 40 tokens per audio second; early fusion preserves the architecture and text-speed training, but output remains text only.
Training moved through unpaired, paired, cooldown and 8k-to-262k context stages; more than half the tokens remained text, using refreshed FineWeb2-HQ+, FinePDFs and Nemotron-CC data. New publicly available, permissively licensed data respects robots.txt opt-outs retroactively. Post-training used 3.7 million SFT samples, full-window SFT, RLVR on about 350,000 prompts, DPO and charter alignment via SDPO, with SDPO only on 8B. Open weights, data, values, recipes, mixtures, intermediate checkpoints and evaluation records accompany the release. The 8B model beats similarly sized open-weight rivals in instruction following and multilingual tasks; thinking lifts AIME, multimodal results approach Gemma 4 12B, Swiss Legal performance matches leading open-weight models, and translation matches GPT-OSS-20B while beating Claude Haiku 4.5 and Google Translate. It still trails frontier open-weight models overall, and speech is experimental.
EvalEval Coalition cards use the Every Eval Ever schema, covering 56 of 60 reporting elements in EU AI Act Article 55 and the GPAI Code of Practice. Proton Lumo and Infomaniak Euria integrated Apertus 1.5; Swisscom, Infomaniak and Phoeniqs announced developer APIs. Hugging Face analytics passed five million downloads. At Zurich’s AI+X Summit, the team previewed Apertus 2.0 with sparse mixture-of-experts architecture, roughly 600B total parameters, longer training, and agent, mathematics and coding priorities.
Positives
- The 262,144-token context window is four times longer than Apertus 1.0 and receives dedicated full-window supervised fine-tuning.
- The 8B model surpasses similarly sized open-weight models on instruction-following and multilingual evaluations.
- Translation performance matches GPT-OSS-20B and exceeds Claude Haiku 4.5 and Google Translate on the Last Translation Benchmark.
- Open weights, datasets, recipes, intermediate checkpoints and evaluation records let researchers reproduce each training stage.
- Proton Lumo and Infomaniak Euria integrated Apertus 1.5, while Swisscom, Infomaniak and Phoeniqs announced developer APIs.
- Five million Hugging Face downloads demonstrate substantial adoption of the Apertus models.
Risks & concerns
- Apertus 1.5 still trails frontier open-weight models overall despite its gains in coding and instruction following.
- Spoken-language processing remains experimental, limiting confidence in the new audio capability.
- Both models understand images and audio but generate only text, restricting their multimodal output.
- Charter alignment through SDPO applies only to 8B, leaving the final 70B model at the DPO stage.
- The Every Eval Ever schema covers 56 of 60 identified EU AI Act and GPAI reporting elements, leaving four unrepresented.