Apertus 1.5: Switzerland's Open AI Turns Multimodal, Context 4x Longer

What Is Apertus 1.5?

On 24 July 2026, ETH Zurich, EPFL and the Swiss National Supercomputing Centre (CSCS) released Apertus 1.5, the second generation of their open large language model. Building on Apertus 1.0 from September 2025, the new release adds native image understanding, an optional thinking mode, a context window quadrupled to 262,144 tokens, and improved instruction-following and tool use. The 8B model received 4 trillion additional training tokens, the 70B model 2 trillion — both remain freely available under the same open licensing approach as Apertus 1.0 (Apache 2.0), fully open and self-hostable.

Four Core Enhancements, Explained

• Native image understanding: documents, technical drawings, diagrams and photos are processed directly, without a separate OCR pipeline.
• Thinking mode: an optional reasoning step before the response improves performance on multi-step, strategic tasks.
• Long context: 262,144 tokens allow entire annual reports, technical manuals or long dialogue histories to be processed without chunking or information loss.
• Tool use and instruction-following: more precise command execution and more robust integration with external tools and APIs — the foundation for agentic workflows.

Why This Matters for Swiss Companies

For Swiss organisations, Apertus 1.5 makes sovereignty practically usable. Data stays on Swiss infrastructure, dependency on US cloud providers is avoided, and the model is built to align with revDSG requirements and the transparency obligations under Article 50 of the EU AI Act.

The Swiss AI Initiative positions Apertus as a cornerstone of a sovereign AI stack, flanked by the Council of Europe AI Framework Convention, the Canton Zurich AI Innovation Programme, and a growing network of European Digital Innovation Hubs (EDIHs).

No More Vendor Lock-In: A Lower Total Cost of Ownership

Because Apertus 1.5 is an open-weight model that can be self-hosted, the classic per-token pricing calculation disappears. For companies with predictable, high request volumes, this fundamentally shifts the cost structure — away from variable API fees and toward one-time infrastructure and operating costs. This dynamic is a driving force behind the industry-wide pricing pressure that proprietary providers such as OpenAI, Anthropic and Google are now facing.

Multimodality in Practice

In the official release material, the team demonstrates the new image processing capability using a technical cross-section drawing of the Niederried-Radelfingen hydropower plant: Apertus 1.5 correctly interprets components, labels and their relationships, describing them in natural language. For Swiss companies, this opens concrete applications in contract analysis, technical documentation, quality control and multilingual customer support — areas often still handled manually or with expensive specialised tools.

Multilingualism and Ecosystem

Apertus 1.0 was already trained on more than 1,000 languages, 40 percent of them non-English, including Swiss German and Romansh. Community projects such as Lightly.ai's Swiss German fine-tuning demonstrate how adaptable the open model is. How Switzerland is building an independent position against US and Chinese AI giants with Apertus is examined in Apertus LLM from ETH Zurich: Europe's Answer to US and Chinese AI Giants.

What This Means for Your AI Strategy

Apertus 1.5 is not a reason to replace an existing setup overnight — but it is a strong argument against binding your architecture to a single model. Companies planning agentic workflows today should factor open alternatives such as Apertus into their architecture from the outset, rather than locking in early.

Conclusion

Apertus 1.5 turns an academic showcase into a production-ready tool: multimodal, long-context, tool-capable and fully open. For Swiss decision-makers, this is the moment to move from discussing sovereign AI to actually evaluating it.