Apertus 1.5: Open Multimodal Models with Longer Context and Improved Reasoning

Apertus 1.5 Release Notes

Apertus 1.5 extends the 8B and 70B Apertus 1.0 models with image and audio understanding, optional thinking mode, a four-times-longer context window, improved instruction-following, and enhanced tool use.

The updated 8B and 70B model weights were released on Hugging Face on July 24, 2026, under the Apache 2.0 license. Further details are provided in the technical report, “Apertus 1.5: From Open Text Models to Open Multimodal Reasoners.”

Apertus 1.5 is a continued pretraining of the first-generation models. The 8B model adds 4 trillion tokens of text and multimodal data, while the 70B model adds 2 trillion tokens. The resulting training totals are 19T tokens for the 8B model and 17T for the 70B model.

The release provides open weights, data documentation, values described in the Apertus Charter, and training details.

Key enhancements

• Native multimodal understanding: The models accept images and audio alongside text, allowing them to process documents, diagrams, and photographs. Spoken-language processing is also supported, although it remains experimental.
• Thinking mode: A selectable mode allows the models to reason about an input before answering, improving results on reasoning tasks.
• Long context: The context window increases to 262,144 tokens, four times the size of the Apertus 1.0 window.
• Improved instruction-following: The models follow instructions more consistently and produce more predictable responses.
• Improved tool use: Training improvements support more effective interaction with external tools and APIs.
• Open values: The models more closely follow the principles described in the Apertus Charter.

Training process

Apertus 1.5 uses four stages of continued pretraining, followed by supervised fine-tuning, long-context fine-tuning, reinforcement learning, preference optimization, and charter alignment.

Multimodality through the tokenizer

Images and audio are introduced through early fusion. They enter the model as discrete tokens in the same sequence as text tokens, rather than through a separate vision encoder.

Images are tokenized with the Emu3.5 tokenizer. A 512×512 image becomes 1,024 tokens. Audio is processed with WavTokenizer at 40 tokens per second. The model architecture remains unchanged because the vocabulary is extended to include the additional modalities. This allows training to run at text-only speed.

Apertus 1.5 understands images and audio but generates text only.

Pretraining

Starting from Apertus 1.0, continued pretraining takes place in four stages:

1. Unpaired data: Text, standalone images, and standalone audio introduce the new modalities.
2. Paired data: Image-caption and audio-transcript pairs align the additional modalities with language.
3. Cooldown: Higher-quality data, including more mathematics and code, is used near the end of pretraining.
4. Long context: The context window is expanded incrementally from 8k to 262k tokens.

More than half of the training tokens are text-only, preserving the language capabilities of Apertus 1.0. Text sources are refreshed with datasets including FineWeb2-HQ+, FinePDFs, and Nemotron-CC.

The new data, including multimodal data, follows the same requirements as earlier releases: it is publicly available, permissively licensed, and respects robots.txt opt-outs retroactively. The sources are documented in the Apertus data-transparency repository.

Post-training

Post-training follows five stages:

SFT → long-context SFT → RL → DPO → charter alignment (SDPO)

1. Supervised fine-tuning: SFT uses 3.7M text samples together with image and speech data. Reasoning traces are included only for prompts the model cannot already solve without them. Tool-use data is limited to execution-verified trajectories.
2. Long-context SFT: Fine-tuning at the full 262k-token window restores long-context capabilities that can be weakened by short-context SFT.
3. RL with verifiable rewards: RLVR uses about 350k mathematics, code, instruction-following, multilingual, and vision prompts, with thinking mode enabled for some examples and disabled for others.
4. DPO: Judge-scored preference pairs improve chat quality, tool use, and safety.
5. SDPO: Self-distillation and judge feedback align the model with the Apertus Charter.

Evaluation results

Evaluation was expanded beyond the text-based tests used for Apertus 1.0. The new evaluation scope includes image and audio understanding, thinking mode, translation, multilingual capabilities, safety, tool use, and additional text benchmarks.

Multimodal evaluations use the MLLM-eval-suite pipeline, which combines the lmms-eval and VLMEvalKit frameworks. Post-training evaluations are available through the project’s evaluation repository, based on a fork of lm-evaluation-harness.

Apertus 1.5 remains behind frontier open-weight models in some areas, but the release improves coding and instruction-following performance. Apertus 1.5 8B outperforms similarly sized open-weight models on instruction-following and multilingual tasks. Thinking mode provides additional gains on difficult reasoning benchmarks such as AIME.

The multimodal performance of Apertus 1.5 is closer to Gemma 4 12B and other current open-weight models. The model is competitive among open-weight models for spatial understanding and among open-source models for document and chart understanding.

On Swiss Legal tasks, Apertus 1.5 shows performance comparable to leading open-weight models. On the Last Translation Benchmark, Apertus 1.5 8B matches GPT-OSS-20B and outperforms models such as Claude Haiku 4.5 and Google Translate.

Evaluation Cards and reproducibility

Apertus 1.5 results are published using the Every Eval Ever (EEE) schema developed with the EvalEval Coalition. The format records scores together with information about who ran the evaluation, how the model was accessed, generation settings, and metric interpretation.

The results are also available through Hugging Face Community Evals. Because the text evaluations use a public fork of lm-evaluation-harness supported by an EEE converter, researchers can rerun the evaluations under identical settings.

Structured evaluation records can also support AI governance. EvalEval mapped Evaluation Cards to requirements including Article 55 of the EU AI Act and the GPAI Code of Practice, finding that the EEE schema can represent 56 of the 60 evaluation-reporting elements identified by those frameworks.

The Apertus project also publishes a Public Summary of Training Content and, voluntarily, a Code of Practice document. Machine-readable evaluation records provide an additional way for regulators, deployers, and researchers to find, interpret, and compare evidence about model performance.

Reproducibility

The technical report documents the training recipes, ablation studies, and data mixtures. Intermediate checkpoints are available in the Apertus 1.5 Hugging Face collection, including base weights, SFT, RLVR, and DPO checkpoints. Curated training data mixtures and custom datasets are available in a separate Apertus 1.5 datasets collection.

The released checkpoints allow researchers to examine how each training stage changes model behavior.

Ablation studies examined the effects of post-training data at different stages. The results indicate that mixing multimodal data with text during SFT improves performance across modalities. Multimodal reasoning traces also continue to provide benefits during later text-only reinforcement learning. Starting from an SFT checkpoint trained with multimodal traces, RL improves mathematics more than it does from a text-only checkpoint, and the reasoning advantage persists during training.

Model checkpoints

Training components

Model availability

Proton Lumo and Infomaniak Euria have integrated Apertus 1.5. Other providers have announced API access, including Swisscom, Infomaniak, and Phoeniqs.

Instructions for running the models are provided in the Hugging Face model cards and the project quickstart guide.

The project has also announced a roadmap for Apertus 2.0. It includes a sparse or mixture-of-experts architecture, approximately 600B total parameters, an extended training schedule, and priorities in agents, mathematics, and coding.

Citation

Project Apertus et al. “Apertus 1.5: From Open Text Models to Open Multimodal Reasoners.” arXiv:2610.15542. Preprint, arXiv, October 11, 2026. https://doi.org/10.48550/arXiv.2610.15542

@misc{apertus2026apertus15opentext,
  title={Apertus 1.5: From Open Text Models to Open Multimodal Reasoners},
  author={Project Apertus and Yixuan Xu and Yassine Turki and Vinko Sabolčec and Valentina Pyatkin and Tommy Chu and Sven Najem-Meyer and Sean F. Carroll and Robert Smith and Raphael Kreft and Oscar Alessandro Tazza and Oleg Lavrovsky and Nicola Irmiger and Matteo Santelmo and Marian Schneider and Juan García Giraldo and Hanna Yukhymenko and Eduard Ďurech and Davit Melikidze and Dani Tamayo and Bettina Messmer and Aryan Ahadinia and Anunay Yadav and Allen Huang and Ahmad Fraij and Alessandro Bombarda and Alex Padula and Alexander Hägele and Alexander Sternfeld and Anna Hedström and Anna Kosovskaia and Arthur Wuhrmann and Arundhati Balasubramaniam and Badr Al Mahouri and Barna Pasztor and Bartosz Szostakiewicz and Bich Ngoc Doan and Bryan Schmid and Burak Sezen and Camille Challier and Clément Charmillot and Dana Shayakhmetova and Daniel Perruchoud and David Brüggemann and David Isztl and Elia Palme and Emma Anastassova and Fabian Bösch and Faruk Zahiragić and Fawzi Roberto Mohamed and Fiona Panicali and Florian Kolly and Flynn Guerra and Francesco Monti and Frederike Lübeck and Gennaro Oliva and Haofeng Yang and Haoliang Shang and Helen El Khoury and Henrique Mendonça and Jannis Vamvas and Jérome Landolt and Jessica Lam and Jingchun Xiao and Jonathan Coles and Joshua Oyebanji and Khanh Nguyen and Konstantinos Pitas and Lasse B. Strand and Li Mi and Loïc Deslarzes and Lorenzo Baggi and Lorenzo Noci and Luca Baumann and Luca Sartori and Lukas Drescher and Marco De Negri and Marco Passerini and Mark Klein and Marko Stojanovic and Martin Majsec and Matin Ansaripour and Mattea Busato and Mauro Pellonara and Max Neuwinger and Michael Graber and Michal Tešnar and Michelle Wastl and Miguel Gila and Mikolaj Boronski and Morgane Magnin and Moritz Reihs and Mukhammadali Sayfiddinov and Nathan Ranchin and Nicolas Cuveillier and Oscar Wohlfahrt and Qianqian Zou and Rada Kamysheva and Rita Sevastjanova and Roman Machacek and Sebastian Keller and Simon Matrenok and Simon Pierre and Stefano Schuppli and Suei-Wen Chen and Szabina Horváth-Mikulas and Taliesin Perez and Thilo Spinner and Thomas Kleine Buening and Tim Robinson and Tymoteusz Kwiecinski and Vilém Zouhar and Vinzenz Thoma and Xiaozhe Yao and Yannick Metz and Yuhei Fukuhara and Yusif Askari and Zixuan Chen and Alexander Ilic and Ana Klimovic and Andreas Krause and Andrei Kucharavy and Elliott Ash and Joost VandeVondele and Martin Rajman and Mennatallah El-Assady and Thomas Schulthess and Timo Kehrer and Torsten Hoefler and Antoine Bosselut and Martin Jaggi and Imanol Schlag},
  year={2026},
  eprint={2610.15542},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2610.15542}
}