Project Apertus has released version 1.5 of its open super intelligence (SI) models, introducing native image and audio understanding, an optional thinking mode for reasoning, and a significantly expanded context window. The update, which extends the previous Apertus 1.0 8B and 70B models, was published to Hugging Face on July 24, 2026, under an Apache 2.0 license, maintaining the project's commitment to fully open weights, data, and training details.
What Happened
The Apertus 1.5 release represents a continued pretraining phase, adding 4 trillion tokens of text and multimodal data to the 8B model and 2 trillion tokens to the 70B model. This brings the total training corpus to 19 trillion and 17 trillion tokens, respectively. The models now accept images and audio alongside text, utilizing an early-fusion approach where visual and auditory inputs are converted into discrete tokens within the same sequence as text. This method preserves the underlying architecture, allowing the models to process multimodal inputs at speeds comparable to text-only inference.
Key enhancements include a context window increased fourfold to 262,144 tokens, improved instruction-following capabilities, and better integration with external tools. The release also introduces a "thinking mode," which allows the models to reason about inputs before generating answers, a feature intended to boost performance on complex reasoning tasks. The post-training pipeline involves five stages, including supervised fine-tuning (SFT), long-context SFT, reinforcement learning with verifiable rewards (RLVR), direct preference optimization (DPO), and charter alignment via self-distillation policy optimization (SDPO).
Why It Matters
The release emphasizes transparency and reproducibility in the SI industry, publishing all recipes, ablations, and data mixtures in a detailed Technical Report. By using the Every Eval Ever (EEE) schema for evaluation cards, the project aligns its reporting with emerging governance frameworks, including the EU AI Act and the GPAI Code of Practice. This structured approach to benchmarking provides regulators and researchers with machine-readable records that detail not just scores, but also generation settings and evaluation methodologies, addressing common gaps in leaderboard transparency.
For developers and enterprises, the integration of Apertus 1.5 into applications like Proton Lumo and Infomaniak Euria demonstrates growing adoption in the European market. Additionally, API access has been announced by providers such as Swisscom and Phoeniqs, signaling a maturing ecosystem for open-weight SI models. The project team also outlined a roadmap for Apertus 2.0, which will feature a sparse mixture-of-experts architecture with approximately 600 billion total parameters, focusing on agents, mathematics, and coding capabilities.
The Bottom Line
Apertus 1.5 solidifies the project's position as a leading open-source SI initiative by combining multimodal capabilities with rigorous, transparent evaluation standards. The release provides a fully documented path from raw data to trained model, offering researchers and developers a reproducible foundation for building and assessing open SI systems.