Apertus

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)

Apertus 1.0 publishes its pretraining-data reconstruction scripts, Megatron-LM training recipe, intermediate checkpoints and technical report under Apache-2.0, as OLMo and Pythia do. Its acceptable-use policy restricts conduct, not commercial use. Apertus 1.5, released 2026-07-24, carries the same license behind an automatically approved click-through, but its training pipelines and checkpoints have not shipped yet, so the entry reflects the still-public 1.0 stack.

• 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://arxiv.org/html/2509.14233 recorded 2026-09-29

  Apertus technical report (HTML rendering): the pretraining codebase at github.com/swiss-ai/Megatron-LM "is built on NVIDIA" Megatron-LM and "We extend the codebase with multiple functionalities"; for data, the authors "release the pipeline code" at github.com/swiss-ai/pretrain-data "to recreate all of the data that was used for training the models".

• https://ethz.ch/en/news-and-events/eth-news/news/2025/09/press-release-apertus-a-fully-open-transparent-multilingual-language-model.html recorded 2026-09-29

  ETH Zurich press release titled "Apertus: a fully open, transparent, multilingual language model", opening "On 2 September 2025, EPFL, ETH Zurich and the Swiss National Supercomputing Centre (CSCS) released Apertus". Corroborates the release framing and establishes no dimension on its own.

• 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/api/models/swiss-ai/Apertus-70B-2509 recorded 2026-09-29

  Hub API record for Apertus-70B-2509 is not gated and not private, carries the license:apache-2.0 tag and 30 safetensors shards, and embeds the "Apertus LLM Acceptable Use Policy" text in its card data, a conduct and indemnity policy with no commercial or scale limit.

• https://huggingface.co/api/models/swiss-ai/Apertus-70B-2509/refs recorded 2026-09-29

  Hub refs listing for Apertus-70B-2509 shows intermediate-checkpoint branches from "step25000-tokens210B" through "step1150000-tokens14920B" plus the final step1155828-tokens15T, and a series of longctx-step branches.

• https://huggingface.co/swiss-ai/Apertus-70B-2509 recorded 2026-09-29

  Apertus-70B-2509 model card, License: apache-2.0, describes a "Fully open model" with "open weights + open data + full training details including all data and training recipes", lists "Training data reconstruction scripts" under open resources, and states "The training intermediate checkpoints are available on the different branches of this same repository". BF16 safetensors shards are listed for download.

• https://huggingface.co/swiss-ai/Apertus-70B-2509/raw/main/LICENSE.txt recorded 2026-09-29

  LICENSE.txt in the Apertus-70B-2509 repo opens "Copyright 2025 - The Swiss AI team." followed by the unmodified "Apache License Version 2.0, January 2004" text.

• https://huggingface.co/swiss-ai/Apertus-v1.5-70B recorded 2026-09-29

  Apertus-v1.5-70B model card, License: apache-2.0, describes a fully open model with "Open weights + open data + open values + full training details including all data and training recipes", the models "handle contexts of up to 262,144 tokens", and "The technical report with further details along with benchmark results, training pipelines, and intermediate checkpoints will be published in the coming weeks". The download now sits behind a click-through: "you accept to share your contact information (email and username) with the repository authors".

Adoption is measured as Hugging Face downloads summed across the family's three shipped checkpoints, most of them now for the multimodal v1.5 release that shipped 2026-07-24. The original 2509 flagship draws little of that traffic today, a shift from when it was the only checkpoint available.

• https://huggingface.co/api/models/swiss-ai/Apertus-70B-2509 recorded 2026-10-07

  32,379 downloads in the trailing 30 days for swiss-ai/Apertus-70B-2509

• https://huggingface.co/api/models/swiss-ai/Apertus-v1.5-70B recorded 2026-10-07

  12,477 downloads in the trailing 30 days for swiss-ai/Apertus-v1.5-70B

• https://huggingface.co/api/models/swiss-ai/Apertus-v1.5-8B recorded 2026-10-07

  697,138 downloads in the trailing 30 days for swiss-ai/Apertus-v1.5-8B

Apertus-70B averages 67.5% on pretraining benchmarks, matching Llama 3.1-70B's 67.3%, and is unusually strong multilingually. Its instruct model trails the newer frontier on reasoning, scoring 69.6% on MMLU against 87.5% for Llama 3.3-70B and 86.6% for Qwen2.5-72B.

• https://arxiv.org/pdf/2509.14233 recorded 2026-08-13

  Apertus technical report, benchmark results and methodology

• 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://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.