Apertus 1.5 Release Notes

The landscape of open-weights artificial intelligence has shifted once again. Project Apertus has officially unveiled Apertus 1.5, a significant evolution of its foundational 8B and 70B models. This release represents a major leap forward, transforming the original text-only architecture into a versatile, multimodal powerhouse capable of processing images, audio, and complex reasoning tasks.

Released on July 24, 2026, under the permissive Apache 2.0 license, the updated model weights are now live on Hugging Face. This update is not merely a refinement; it is a comprehensive expansion that incorporates 4 trillion additional tokens for the 8B model and 2 trillion for the 70B variant, bringing the total training corpus to 19 trillion and 17 trillion tokens, respectively.

Key Enhancements: A New Standard for Open Models

Apertus 1.5 introduces a suite of capabilities designed to bridge the gap between specialized research models and practical, real-world utility. The development team has focused on six core pillars of improvement:

• Native Multimodal Understanding: The models now ingest images and audio alongside text, allowing for seamless interaction with diagrams, documents, and visual media. Experimental support for spoken language processing is also included.
• Thinking Mode: Users can now toggle a dedicated reasoning mode. By allowing the model to "think" before generating a final response, performance on complex logic and math problems is substantially boosted.
• Expanded Context Window: The context capacity has been quadrupled, now supporting 262,144 tokens, enabling the analysis of massive documents and long-form codebases.
• Refined Instruction Following: The models demonstrate a higher degree of predictability and adherence to complex user prompts compared to the 1.0 generation.
• Advanced Tool Integration: Enhanced training protocols allow Apertus 1.5 to interact more effectively with external APIs and software tools.
• Open Values: The release reinforces the Apertus Charter, ensuring that the model’s principles remain transparent and aligned with the project’s core mission.

"As with every Apertus release, the models are fully open: open weights, open data, open values, and full training details. We believe that true progress in AI requires radical transparency."

The Architecture of Multimodality

A departure from traditional vision-encoder architectures, Apertus 1.5 utilizes an early-fusion approach. By treating images and audio as discrete tokens within the same sequence as text, the model processes all modalities through a unified pipeline.

Images are tokenized using the Emu3.5 tokenizer, where a standard 512×512 image is converted into 1,024 tokens. Audio is processed via WavTokenizer at a rate of 40 tokens per second. Because the team opted to extend the vocabulary rather than modify the underlying architecture, the model maintains high-speed performance, running at speeds comparable to text-only models.

The Four-Stage Pre-Training Pipeline

The transition to a multimodal, long-context model involved a rigorous four-stage pre-training process: 1. Unpaired Training: Exposure to standalone text, images, and audio to build foundational modality recognition. 2. Paired Training: Alignment of image-caption and audio-transcript pairs to bridge the gap between sensory data and language. 3. Cooldown: A final phase utilizing high-quality math and code datasets to sharpen precision. 4. Long-Context Scaling: A step-by-step expansion of the context window from 8k to 262k tokens.

Throughout this process, over 50% of the training tokens remained text-only, ensuring that the linguistic proficiency established in Apertus 1.0 was not only preserved but enhanced.

Post-Training and Alignment

The post-training phase is where Apertus 1.5 truly finds its voice. The team employed a five-stage strategy: SFT (Supervised Fine-Tuning), Long-Context SFT, Reinforcement Learning (RL), Direct Preference Optimization (DPO), and Charter Alignment (SDPO).

A critical innovation here was the use of RLVR (Reinforcement Learning with Verifiable Rewards). By training on approximately 350,000 prompts—spanning math, coding, and multilingual tasks—the model learned to balance "thinking" and "non-thinking" responses effectively. Furthermore, the use of SDPO (Self-Distillation for Preference Optimization) ensures that the model’s outputs remain strictly aligned with the Apertus Charter, providing a safety layer that is both robust and transparent.

Benchmarking and Evaluation

Apertus 1.5 has undergone an extensive evaluation battery. Beyond standard text benchmarks, the team integrated the MLLM-eval-suite, which merges the lmms-eval and VLMEvalKit frameworks to assess multimodal performance.

The results are promising. Apertus 1.5 8B currently outperforms similarly sized open-weight models in instruction following and multilingual tasks. In the realm of coding and spatial reasoning, the model shows significant gains, with multimodal performance now rivaling established models like the Gemma 4 12B. Notably, the model excels in Swiss Legal tasks and the Last Translation Benchmark, where it has demonstrated capabilities that exceed those of Claude Haiku 4.5.

The EvalEval Coalition and Transparency

In a move to standardize how AI performance is reported, Project Apertus has partnered with the EvalEval Coalition. They are publishing results using the Every Eval Ever (EEE) schema. Unlike traditional leaderboards that often hide the "how" behind a score, EEE records:

• The identity of the evaluator.
• The specific generation settings used.
• The environment and access method.
• Detailed metric interpretations.

This approach aligns with emerging global standards, including Article 55 of the EU AI Act. By making these records machine-readable, Apertus is setting a new bar for accountability in the AI industry.

Reproducibility and Community Access

Project Apertus remains committed to the idea that research is only as good as its reproducibility. The team has hosted all intermediate checkpoints—from base weights to final SFT and DPO models—on their Hugging Face collection.

Researchers can now explore the "why" behind the model’s behavior by examining the specific training data mixes and ablations documented in the technical report. The team noted that mixing multimodal data during the SFT stage provided a "lift" in performance across all modalities, a finding that should prove invaluable to other developers in the open-source space.

Deployment and Future Roadmap

Apertus 1.5 is already seeing real-world adoption. Applications such as Proton Lumo and Infomaniak Euria have integrated the model, and API access is being provided by partners including Swisscom, Infomaniak, and Phoeniqs.

Looking ahead, the team is already focused on the future. During the recent AI+X Summit in Zurich, the roadmap for Apertus 2.0 was revealed. The next generation will move toward a sparse/MoE (Mixture of Experts) architecture with a massive 600B total parameter count, with a specific focus on agentic workflows, advanced mathematics, and high-level coding capabilities.

Conclusion

Apertus 1.5 is more than a version update; it is a testament to the viability of the open-source model development philosophy. By combining massive scale with granular transparency and a commitment to multimodal reasoning, the team has provided the community with a powerful tool for innovation.

For those looking to get involved, the project encourages developers to download the weights, explore the datasets, and contribute to the ongoing evaluation efforts. As the community continues to grow—having recently surpassed five million downloads—Apertus stands as a beacon for what is possible when the barriers to AI research are dismantled.

*

Citation: Project Apertus, et al. “Apertus 1.5: From Open Text Models to Open Multimodal Reasoners.” arXiv:2610.15542. Preprint, arXiv, October 11, 2026.