TL;DR

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Apertus 1.5, released on July 24, 2026, extends previous models with multimodal input, a 262,144-token context window, and improved reasoning and tool use. The open-source models now support images and audio alongside text, with significant performance gains.

The Apertus project announced the release of Apertus 1.5 on July 24, 2026, introducing major enhancements including native multimodal understanding, a fourfold increase in context length, and improved instruction-following capabilities. The models, based on the 8B and 70B architectures, are now capable of processing images and audio alongside text, significantly broadening their application scope and usability for open-source AI development.

Apertus 1.5 extends the capabilities of the original Apertus models through extensive pretraining, adding 4 trillion tokens of multimodal data for the 8B model and 2 trillion for the 70B model, reaching a total of 19 trillion and 17 trillion tokens respectively. The models were released to Hugging Face under an Apache 2.0 license, with full transparency on training data and methodology, including open weights and data sources.

The key innovations include native multimodal understanding, achieved by integrating images and audio as tokens within the same sequence as text, using the Emu3.5 tokenizer for images and WavTokenizer for audio. The models can interpret documents, diagrams, photos, and spoken language (though speech remains experimental). The context window has increased from 8,000 to 262,144 tokens, enabling the models to handle much longer inputs, which improves performance on complex reasoning tasks.

Additional improvements focus on instruction-following, tool use, and alignment with the Apertus Charter, emphasizing transparency and adherence to defined values. The training process involved multiple stages, including unpaired multimodal pretraining, paired data alignment, and fine-tuning with reinforcement learning and self-distillation, aiming to enhance safety, accuracy, and utility. Evaluation results indicate that Apertus 1.5 outperforms previous models on various benchmarks, including multilingual translation, legal tasks, and reasoning-heavy tests such as AIME, bringing it closer to current frontier open-weight models.

Implications of Apertus 1.5 for Open-Source AI

The release of Apertus 1.5 marks a significant milestone in open-source AI development, especially in multimodal understanding and long-context reasoning. By making these capabilities accessible in fully open models, the project democratizes advanced AI tools, enabling broader experimentation, customization, and deployment outside proprietary systems. The improvements in instruction-following and tool use also enhance the models’ practical utility for developers and researchers.

This development could accelerate research in multimodal AI, foster transparency, and challenge proprietary models by demonstrating that open models can achieve competitive performance across complex tasks. However, the open nature also raises questions about safety, misuse, and ethical considerations, which the Apertus team addresses through rigorous evaluation and alignment processes.

Background and Development of Apertus Models

The Apertus project has been evolving since its initial release, focusing on open, transparent AI models trained on large, publicly available datasets. The previous version, Apertus 1.0, demonstrated strong language understanding but lacked multimodal capabilities and had limited context length. The new release builds on this foundation by integrating multimodal data and expanding the context window, driven by advances in tokenizer technology and training methodologies.

Prior to Apertus 1.5, open-weight models generally lagged behind proprietary systems in multimodal and reasoning tasks. The Apertus team emphasized a multi-stage training pipeline, combining unpaired multimodal pretraining, paired data alignment, and reinforcement learning, to improve performance and safety. The release aligns with ongoing industry trends toward larger, more capable open models and increased transparency in AI development.

Open data sources like FineWeb2-HQ+ and Nemotron-CC, along with rigorous evaluation protocols, underpin the transparency and reproducibility of Apertus 1.5, setting a new benchmark for open-source AI projects.

Remaining Questions About Apertus 1.5 Capabilities

While Apertus 1.5 demonstrates impressive performance, some aspects remain uncertain. Speech understanding, for example, is still experimental, and real-world performance in diverse applications needs further validation. The models’ safety, robustness, and potential for misuse are also ongoing concerns, despite the project’s emphasis on alignment and transparency. Additionally, the long-term impact of such large models on the open-source ecosystem and AI ethics remains to be seen.

Future Developments and Next Steps for Apertus

The Apertus team plans to continue refining the models through additional training stages, including more advanced safety measures and domain-specific fine-tuning. They will also expand evaluation efforts, especially in real-world scenarios, and seek community feedback to guide future improvements. A key milestone will be the release of more extensive documentation and tools to facilitate broader adoption and experimentation by researchers and developers.

Further integration of multimodal capabilities into practical applications, such as document analysis, creative tools, and interactive systems, is expected to follow. The team also hints at exploring speech recognition improvements and more sophisticated reasoning modules to enhance the models’ versatility and safety.

Where I land

In my view, Apertus 1.5 marks a significant step in democratizing advanced AI technology, especially with its multimodal capabilities and open approach. However, I remain cautious about the safety and ethical implications of deploying such powerful models broadly. The open-source nature accelerates innovation but also raises risks of misuse or unintended consequences. What could change my assessment is clear evidence of effective safety measures and real-world testing that demonstrate responsible use. Until then, I see this as an exciting but cautious development in the open AI landscape.

Key Questions

What are the main new features in Apertus 1.5?

The main new features include native multimodal understanding (images and audio), a 262,144-token context window, improved instruction-following, and better tool integration.

Is Apertus 1.5 open source?

Yes, all weights, data, and training details are fully open, licensed under Apache 2.0, promoting transparency and community collaboration.

How does Apertus 1.5 perform compared to other models?

It shows competitive performance in multilingual translation, reasoning, and legal tasks, with improvements over previous Apertus versions and close to some frontier open-weight models in multimodal understanding.

What are the limitations of Apertus 1.5?

Speech understanding remains experimental, and safety, robustness, and misuse prevention are ongoing challenges. Real-world deployment and long-term impacts are still under assessment.

What are the plans for future updates?

The team intends to improve safety, expand multimodal capabilities, enhance reasoning, and release more tools and documentation to foster wider adoption and experimentation.

Source: Hugging Face