Apertus 1.5: Switzerland’s Open‑Source AI Model Takes a Leap
Key takeaways
- Apertus 1.5 scales to 13 billion parameters and reduces inference latency by ~30 % thanks to sparse‑attention kernels.
- The model now supports a broader set of European languages, improving multilingual performance on benchmarks.
- Built‑in, open‑source toxicity filters align the model with the EU AI Act, enabling transparent responsible‑AI practices.
- Training leveraged CSCS’s Piz Daint‑X supercomputer, demonstrating that public infrastructure can deliver world‑class AI performance with a modest carbon footprint.
- Open‑source release under Apache 2.0 and Hugging‑Face compatibility encourages rapid adoption across academia, industry, and public services.
Introduction
The AI landscape is dominated by a handful of commercial giants, yet Europe continues to nurture home‑grown alternatives that prioritize openness, reproducibility, and ethical safeguards. The most recent milestone in this effort is Apertus 1.5, the latest iteration of Switzerland’s open‑source large language model (LLM). Launched by the Swiss National Supercomputing Centre (CSCS) in collaboration with several academic partners, Apertus 1.5 demonstrates how a publicly funded ecosystem can keep pace with the rapid advances of private sector models while staying true to the principles of transparency and data sovereignty.
What Is Apertus?
Apertus, named after the Latin word for “open,” was first introduced in 2023 as a 7‑billion‑parameter transformer trained on a curated mix of multilingual European text, scientific literature, and publicly available web data. Its design philosophy mirrors that of other community‑driven projects such as LLaMA and Mistral, but with a distinct emphasis on:
* Data provenance – every document in the training set is traceable to a public licence or an explicit permission. * Energy efficiency – the model was trained on CSCS’s Piz Daint and later on the new Piz Daint‑X GPU cluster, leveraging waste‑heat recovery and carbon‑offset schemes. * Modular tooling – the codebase is hosted on GitHub under an Apache 2.0 licence, allowing researchers to plug in custom tokenisers, adapters, or quantisation pipelines.
The Leap to 1.5
Apertus 1.5 arrives with three headline improvements:
1. Scale and Speed – The parameter count rises to 13 billion, and the model benefits from a revamped sparse‑attention kernel that reduces inference latency by roughly 30 % on NVIDIA H100 and AMD MI250 GPUs. 2. Multilingual Depth – Training data now includes expanded corpora in French, Italian, Romansh, and several Eastern European languages, raising the average BLEU score on the Europarl benchmark from 28.7 to 34.2. 3. Responsible AI Guardrails – A built‑in toxicity filter, calibrated on the European Union’s AI Act guidelines, can be toggled at inference time. The filter is open‑source as well, enabling third‑party audits.
These upgrades were made possible by a combination of hardware upgrades, algorithmic refinements, and a broader community contribution model.
Behind the Scenes: How CSCS Made It Happen
Supercomputing Muscle
The training run spanned 45 days on the Piz Daint‑X system, a hybrid architecture that couples 512 GPU nodes (each equipped with 8 H100s) with a high‑speed InfiniBand network. By exploiting mixed‑precision training (FP8 where supported) and gradient checkpointing, the team kept the total energy consumption under 1.2 GWh – a figure comparable to a small Swiss town’s annual electricity use.
Academic Partnerships
Key research groups from ETH Zurich, the University of Zurich, and the Swiss Federal Institute of Technology Lausanne (EPFL) contributed domain‑specific datasets, particularly in law, medicine, and climate science. Their involvement ensured that Apertus 1.5 performs well on specialized queries, a niche where many commercial LLMs still rely on costly fine‑tuning services.
Open‑Source Ecosystem Integration
Apertus 1.5 ships with a Hugging Face‑compatible model card and a Dockerised inference server. The repository also includes a Ray‑based serving layer that can auto‑scale across on‑premise clusters or public clouds, giving users flexibility while preserving data locality.
Why Apertus 1.5 Matters for Europe
1. Strategic Independence – By owning the full stack—from data collection to model serving—Switzerland reduces reliance on foreign AI providers, aligning with EU policy goals around digital sovereignty. 2. Research Acceleration – Open access to a state‑of‑the‑art LLM lowers the barrier for Swiss and European scholars to experiment with generative AI in fields ranging from computational linguistics to drug discovery. 3. Transparency & Trust – The public training logs, model weights, and evaluation scripts enable independent verification of claims about bias, robustness, and carbon footprint.
Real‑World Use Cases Already Emerging
* Legal Drafting Assistant – A pilot at the Federal Supreme Court of Switzerland uses Apertus 1.5 to suggest clause language in bilingual contracts, cutting drafting time by 20 %. * Scientific Summarisation – Researchers at the Swiss Institute of Bioinformatics have integrated the model into a pipeline that auto‑generates abstracts for pre‑prints, improving discoverability of emerging research. * Multilingual Customer Support – A Swiss fintech startup deployed the model on its help‑desk platform, offering instant, GDPR‑compliant responses in German, French, and Italian.
Challenges and the Road Ahead
While Apertus 1.5 marks a significant step forward, several hurdles remain:
* Data Licensing – Even with diligent curation, the sheer volume of web‑scraped text makes exhaustive licence verification difficult. Ongoing collaboration with the European Data Portal aims to develop a standardized metadata schema. * Model Size vs. Accessibility – At 13 B parameters, the model still requires substantial hardware for low‑latency inference. Future work will explore quantisation to 4‑bit formats without sacrificing accuracy. * Governance – Managing contributions from a global community while maintaining a coherent release schedule calls for a robust governance model. CSCS is drafting a charter that outlines contributor rights, code review processes, and security audits.
Conclusion
Apertus 1.5 showcases that open‑source AI can be both cutting‑edge and responsibly engineered. By leveraging national supercomputing assets, fostering cross‑institutional collaboration, and embedding ethical safeguards from the ground up, Switzerland sets an example for how Europe can cultivate its own AI ecosystem. As the model matures and the community around it expands, we can expect a cascade of innovative applications that respect both user privacy and regional regulatory frameworks.
Stay tuned for the upcoming Apertus 2.0 roadmap, which promises multimodal capabilities and tighter integration with edge devices.