Apertus 1.5: 262K context, image and audio support

What changed

ETH Zurich, EPFL, and the Swiss National Supercomputing Centre released Apertus 1.5 on July 24, 2026, under an Apache 2.0 license. Rather than a complete rebuild, the updated 8B and 70B models extend Apertus 1.0 through continued pretraining. The 8B model received 4 trillion additional tokens, while the 70B model received 2 trillion tokens, bringing cumulative training volumes to 19 trillion and 17 trillion tokens respectively.

The release introduces four main improvements. The models now accept images and audio inputs alongside text through early-fusion tokenization, where images convert to 1,024 tokens via the Emu3.5 tokenizer and audio processes at 40 tokens per second with WavTokenizer. The context window expands to 262,144 tokens from the prior 8,000-token limit. An optional thinking mode allows models to reason before answering, improving performance on reasoning-focused tasks. Enhanced tool use training enables more effective integration with external tools and APIs through execution-verified training trajectories.

All components remain publicly available, including model weights, training data, and development details. The team published intermediate checkpoints, curated training data combinations, and ablation studies in a Technical Report. Training data derives from publicly available sources including FineWeb2-HQ+, FinePDFs, and Nemotron-CC, with retroactive respect for robots.txt opt-outs.

Post-training proceeded through five sequential stages. Supervised fine-tuning covered 3.7 million text samples plus multimodal data. Long-context fine-tuning operated at the full 262K window to restore capabilities that short-context training erodes. Reinforcement learning with verifiable rewards applied to roughly 350,000 samples covering math, code, instruction-following, multilingual, and vision tasks. Direct preference optimization used judge-scored pairs to improve chat quality and tool use. Final alignment to the Apertus Charter occurred through structured direct preference optimization using judge feedback and self-distillation.

Why it matters

This release addresses a practical gap for teams deploying open models in production. Many workflows require analyzing documents, diagrams, and photographs alongside text processing. The multimodal capability handles this without requiring separate vision systems. The quadrupled context window enables longer document processing, extended code repositories, and deeper conversational histories without truncation.

The optional thinking mode targets specific reasoning tasks where intermediate steps improve accuracy. The developers report improvements on reasoning benchmarks like AIME, though performance remains behind frontier proprietary models overall. The benefits vary by task type, making evaluation essential before adoption.

For organizations with local deployment requirements, the full transparency of training methodology and data sourcing enables auditing and compliance assessment. This matters under Article 55 of the EU AI Act, where model documentation and evaluation records carry regulatory weight. The team publishes evaluation results in the Every Eval Ever schema format, which aligns with standards used by the UK AI Security Institute, improving comparability against other models.

The release reflects sustained institutional support from Swiss research organizations, ensuring continuity. However, incremental performance gains on most benchmarks compared to larger frontier models set realistic expectations about capabilities.

What to test

Before deploying Apertus 1.5, evaluate performance on your specific use cases:

• Multimodal performance on your document types. The developers report competitive performance with Gemma 4 12B on spatial understanding and document understanding among open-source models, but note gaps versus the largest proprietary systems. Test charts, PDFs, and diagrams relevant to your domain to assess practical utility.

• Reasoning mode trade-offs. Thinking mode adds computational cost while improving accuracy on certain tasks. Benchmark both latency and accuracy with and without reasoning enabled to understand the cost-benefit for your workload.

• Long-context quality near boundaries. While the 262K window represents genuine expansion, document quality may degrade near token limits. The developers describe a dedicated long-context fine-tuning stage to restore abilities that short-context training damages, but test whether retention remains consistent throughout the full context span.

• Tool-use execution reliability. Training emphasizes execution-verified tool trajectories, which is stricter than some alternatives. Verify that tool calls match your API schemas and handle errors according to your requirements.

• Multilingual performance. Extended translation benchmarking shows the 8B model matches GPT-OSS-20B and exceeds models like Claude Haiku 4.5 and Google Translate on the Last Translation Benchmark. Evaluate your specific language pairs against baseline systems.

• Instruction-following consistency. The release claims improvements over version 1.0 in following user directions. Sample outputs across varied prompt phrasings to confirm consistency meets your tolerance for deviation.

Common questions

Who developed Apertus 1.5?

ETH Zurich, EPFL, and the Swiss National Supercomputing Centre developed Apertus 1.5 as part of a broader Swiss AI Initiative. The models are open-weight and released under Apache 2.0.

How much larger is the context window in Apertus 1.5?

The context window grows to 262,144 tokens from Apertus 1.0’s 8,000-token limit, representing a 4x increase. Long-context fine-tuning at full size helps preserve model quality across the extended range.

Can Apertus 1.5 generate images or audio?

No. Apertus 1.5 accepts images and audio as input for understanding but generates text only. The developers note that audio processing remains experimental.

The conclusion

Apertus 1.5 demonstrates methodical progress in open-weight language models. The multimodal support and expanded context window address concrete deployment needs for specific workflows. Thinking mode provides a reasoning option for tasks where step-by-step problem-solving improves results, though gains depend on task characteristics.

The transparent training pipeline and structured evaluation reporting provide the most value for organizations where auditability matters alongside raw performance. Regulators, enterprises, and researchers can inspect data sources, training stages, and evaluation methods more thoroughly than with closed models. This approach positions Apertus as a reference for responsible model release practices.

Performance trails frontier closed models by design. The 8B variant shows competitive results on instruction-following and multilingual translation among similarly sized open-weight models. The 70B variant targets researchers and organizations operating local infrastructure for increased capability.

Watch adoption rates among European research institutions and regulated industries where open foundations and AI Act alignment influence technology choices. Strong performance on Swiss Legal tasks and multilingual translation suggests the team understands regional requirements. The practical test will be whether multimodal capabilities see production adoption or remain primarily experimental.

Related reading

• Cloudflare releases Clef-omni with audio and video support
• OrcaCyber Zero 1.5 Cybersecurity Model Launches With 1M Context
• Liquid AI Releases Multimodal Open D1 Decision Models for Edge