Apertus
Swiss AI InitiativeSwitzerland's national open LLM family from the Swiss AI Initiative at EPFL, ETH Zurich and CSCS. The current 1.5 models at 8B and 70B take image and audio input alongside text, run a deliberation mode and handle a 262,144-token context. The line covers 1,811 languages, about 40 percent of them non-English, including Swiss German and Romansh.
Apertus 1.5 shipped 2026-07-24 by continued pretraining of 1.0. Its own pipelines and checkpoints are announced but not yet shipped, so openness rests on the still-public 1.0 stack and its data-reconstruction scripts rather than a released corpus. Verified 2026-08-14 via the v1.5 model cards and the pretrain-data repository.
Openness
5 high confidence- weights
- open(Apache-2.0, BF16 safetensors)
- data
- open(full pretraining-data reconstruction scripts, github.com/swiss-ai/pretrain-data)
- code
- open(Megatron-LM training recipe + reconstruction scripts)
- checkpoints
- open(intermediate checkpoints on HF branches)
- license
- Apache-2.0(OSI)
The Apertus family's current release is the multimodal Apertus 1.5, shipped 2026-07-24. The 1.0/2509 release remains public with its full open stack: pretraining-data reconstruction scripts (respecting opt-out and robots.txt and stripping personal data), the Megatron-LM training recipe, intermediate checkpoints, and the technical report (arXiv 2509.14233), all Apache-2.0, which is the OLMo and Pythia full-openness tier. Apertus 1.5 is likewise Apache-2.0 and declares a 'fully open model: open weights + open data + open values + full training details including all data and training recipes', but its own multimodal-mix reconstruction scripts, training pipelines and intermediate checkpoints are announced as coming in the weeks ahead and have not shipped. The 5 therefore rests on the family's established practice and the still-public 1.0 stack rather than on the 1.5 promises. An Acceptable Use Policy accompanies the Apache-2.0 grant, restricting conduct (illegal or abusive use) rather than commerce or scale, so the license stays OSI.
- https://api.github.com/repos/swiss-ai/pretrain-data recorded 2026-08-14
`private: false`, `archived: false`, license `Apache-2.0`, description "Pretraining data reconstruction scripts for Apertus". The reconstruction scripts are published, which is what satisfies data:open by reconstruction rather than by a released corpus.
- https://huggingface.co/swiss-ai/Apertus-70B-2509 recorded 2026-08-14
Model card: "Fully open model: open weights + open data + full training details including all data and training recipes"; "Training Framework: Megatron-LM" linking github.com/swiss-ai/Megatron-LM; "The training intermediate checkpoints are available on the different branches of this same repository"; license apache-2.0; BF16 safetensors.
- https://ethz.ch/en/news-and-events/eth-news/news/2025/09/press-release-apertus-a-fully-open-transparent-multilingual-language-model.html recorded 2026-08-14
ETH Zurich press release, title "Apertus: a fully open, transparent, multilingual language model", dated 2 September 2025 (EPFL/ETH/CSCS). Corroborates the release framing; establishes no dimension on its own, since a press release describes rather than exhibits the artifacts.
- https://github.com/swiss-ai/pretrain-data recorded 2026-08-14
Repository page for the pretraining-data reconstruction scripts; "reconstruction" appears throughout the tree and README.
- https://huggingface.co/swiss-ai/Apertus-v1.5-70B recorded 2026-08-14
Apertus-v1.5-70B model card: "Fully Open Model: Open weights + open data + open values + full training details including all data and training recipes"; License: apache-2.0; identifies itself as "Apertus 1.5 Omni, a multimodal assistant ... you understand images and audio and respond in text"; "context length up to 262,144 tokens, a four-fold increase from our initial Apertus 1.0 release"; and notes that "benchmark results, training pipelines, and intermediate checkpoints will be published in the coming weeks". Establishes weights and license for the current release; data/code for the v1.5 multimodal mix are declared but not yet shipped.
Adoption
2 medium confidence17,574 downloads in the trailing 30 days summed across the family's shipped SKUs (swiss-ai/Apertus-v1.5-8B 11,501; Apertus-v1.5-70B 5,520; Apertus-70B-2509 553). Bands at level 2 (10K-100K) on the model adoption scale, up from level 1 as the multimodal v1.5 release (2026-07-24) drew the bulk of the downloads, while the original 2509 flagship now sees little traffic.
- https://huggingface.co/api/models/swiss-ai/Apertus-v1.5-8B recorded 2026-08-14
`downloads` (trailing 30 days) = 11501 for swiss-ai/Apertus-v1.5-8B; `pipeline_tag` image-text-to-text
- https://huggingface.co/api/models/swiss-ai/Apertus-v1.5-70B recorded 2026-08-14
`downloads` (trailing 30 days) = 5520 for swiss-ai/Apertus-v1.5-70B; `pipeline_tag` image-text-to-text
- https://huggingface.co/api/models/swiss-ai/Apertus-70B-2509 recorded 2026-08-14
`downloads` (trailing 30 days) = 553 for swiss-ai/Apertus-70B-2509
Capability
3 medium confidenceMid-tier capability: competitive with the prior open generation, around the Llama 3.1-70B class, on pretraining benchmarks and unusually strong multilingually, but below the 2026 frontier on MMLU-style reasoning and knowledge. The value proposition is full openness, multilingual breadth and legal compliance rather than raw SOTA - the same profile as OLMo 2. Apertus 1.5 adds multimodal input (image and audio), a thinking mode and a 262,144-token context, and claims performance 'comparable to other models of similar size', but no quantified 1.5 benchmarks have been published yet, so the score rests on the last citable 2509 numbers. Apertus sits level with the Llama family, both at 3, even though Llama's own reading is now taken on Llama 4 Maverick rather than Llama 3.1 405B.
- https://huggingface.co/swiss-ai/Apertus-70B-2509 recorded 2026-08-13
pretraining benchmark table: Apertus-70B avg 67.5% vs Llama 3.1-70B 67.3%; Apertus-70B-Instruct MMLU 69.6%
- https://arxiv.org/pdf/2509.14233 recorded 2026-08-13
Apertus technical report, benchmark results and methodology
- https://huggingface.co/swiss-ai/Apertus-v1.5-70B recorded 2026-08-14
Apertus-v1.5-70B card: multimodal ("you understand images and audio and respond in text"), a deliberation/thinking mode, and "context length up to 262,144 tokens, a four-fold increase"; corroborates the v1.5 capability additions, with quantified benchmarks pending the report.
Verified 2026-08-13