Apertus AI is one of the most transparent and technically mature efforts to build a fully open-source large language model. Developed in Switzerland and released together with its source code, training documentation and model weights, it offers an unprecedented level of reproducibility and independence from closed ecosystems. This makes it ideal for researchers, public-sector institutions and organizations seeking control over their AI infrastructure.
The model comes in two sizes, 8B and 70B parameters, supporting both lightweight and high-performance workloads. Its multilingual focus, combined with detailed technical disclosures, provides a foundation for reliable and responsible AI development.
Comparison with other open LLMs
| Model | Fully Open | Languages | Sizes | License | Highlights |
|---|---|---|---|---|---|
| Apertus | Yes | Multilingual | 8B, 70B | Open | Transparency, reproducibility |
| Llama 3 | Not fully | 30+ | 8B, 70B | Custom | High performance |
| Mistral | Partially | EN, FR | 7B, 8x22B | Apache 2 | Efficiency |
| Phi-3 | Partially | EN | 4B, 7B | MIT | Lightweight |
| Apertus Instruct | Yes | Multilingual | 8B, 70B | Open | Dialogue-ready |
Apertus is among the few models that are truly open in every layer, including access to training descriptions and architectural details.
Ways to integrate Apertus
Developers can download the model from Hugging Face and use it directly through the transformers ecosystem, building chatbots, translation pipelines or domain-specific applications. For production-grade infrastructure, Amazon SageMaker provides cloud deployment with autoscaling, secure endpoints and GPU support. Teams can fine-tune the model using LoRA techniques, especially on the 8B version, which performs well even on mid-range hardware.
Step-by-step installation guide
Environment setup
Install Python 3.9 or later and create a virtual environment.
Install PyTorch, transformers, tokenizers and numpy.
If using GPU, install the appropriate CUDA and cuDNN versions.
Download the model
Visit: https://huggingface.co/swiss-ai/Apertus-8B-2509
Download weights, tokenizer and configuration files.
Local configuration
Load the tokenizer and model.
Select the correct backend (CPU, GPU or MPS for macOS).
MacOS users may opt for mlx-lm to simplify execution.
Optional optimization
Apply quantization or use inference-optimized runtimes for lower memory usage.
Code example resources
Official Apertus documentation: https://huggingface.co/docs/transformers/en/model_doc/apertus
Official fine-tuning recipes: https://github.com/swiss-ai/apertus-finetuning-recipes
Python usage examples: https://skywork.ai/blog/models/swiss-ai-apertus-8b-2509-free-chat-online-skywork-ai/
vLLM integration: https://docs.vllm.ai/en/stable/api/vllm/model_executor/models/apertus/
Apertus combines transparency, multilingual capabilities and deployment flexibility, making it one of the most robust open-source LLMs currently available. It represents a path toward sovereign, trustworthy and accessible AI for organizations and institutions worldwide.
—