What is open-source AI?

Open-source AI vs. open weights - the key distinction

The term open-source AI is used loosely. Strictly, open source means the licence lets anyone use, study, modify and redistribute the software for any purpose. In 2024 the Open Source Initiative published its Open Source AI Definition (OSAID 1.0), which additionally expects enough information about the training data to rebuild a comparable model. Most well-known 'open' language models do not clear that bar.

The more precise term for most released models is open-weight: the trained parameters (weights) are downloadable, so you can run and fine-tune the model yourself, while the training data, full training code or licence terms may stay restricted. Open weights still deliver most practical benefits - local deployment, customization, no per-token fee - even when a model is not 'open source' in the strict sense.

Understanding licences: what matters

• Permissive licences (Apache 2.0, MIT): commercial use allowed; Apache 2.0 also includes a patent grant. Used by Mistral 7B, Mixtral, many Qwen models and Apertus.
• Community/custom licences (Llama Community License, Gemma Terms): commercial use usually allowed but with conditions - acceptable-use rules, naming/branding requirements or thresholds (Llama requires a separate licence above 700M monthly active users).
• Non-commercial / research-only licences: weights available for research, but no commercial production use permitted.
• Check before production: is commercial use permitted? May fine-tuned weights be redistributed? Are there attribution or naming duties? Any output restrictions? Always read the actual licence.

Examples: Llama, Mistral and other open models

• Llama (Meta): large open-weight family under a community licence, with a very large ecosystem and many fine-tuned derivatives - not officially OSI-approved.
• Mistral (France): mixes genuinely open Apache-2.0 models (Mistral 7B, Mixtral 8x7B) with flagship models under commercial licences.
• Gemma (Google): compact open-weight models under custom terms; Qwen (Alibaba): many Apache-2.0 models with strong multilingual ability; DeepSeek: open weights, including notable reasoning models.
• Apertus (Switzerland): a fully open language model from ETH Zurich, EPFL and the CSCS national supercomputing centre, released in several sizes. Weights, training data and training details are disclosed, licensed under Apache 2.0, multilingual - a reference example of genuine openness.

Benefits - especially from a Swiss perspective

• Data sovereignty and privacy: run on-premises or in a Swiss cloud so data stays under your control - helpful for revDSG/nFADP compliance and professional secrecy.
• No vendor lock-in: you are not tied to one provider's pricing or roadmap and can switch models or self-host.
• Cost control at scale and customization: fine-tune on your own domain and language data instead of paying per token.
• Transparency, auditability and longevity: open models can be inspected and cannot be suddenly deprecated out from under you.

Caveats and limits

• Responsibility sits with you: safety guardrails, misuse protection and compliance - including possible EU AI Act duties for providers of general-purpose AI (GPAI) models.
• Infrastructure and expertise: GPUs, MLOps and operations cost money and require skills; 'free to download' does not mean 'free to run'.
• 'Open' is often only partial: licence restrictions, undisclosed training data and use clauses can limit deployment scenarios.
• Security and maintenance: you manage updates, model provenance and supply chains yourself; support is community-based unless you buy a commercial contract. On the hardest tasks, open models can still trail leading closed models.

Open-source AI in Switzerland: when it makes sense

With Apertus, Switzerland shows that digital sovereignty and open AI can go together - an important signal for government, healthcare, banking and SMEs that must keep data in-country. Open models fit especially where privacy, professional secrecy or cost control dominate and you want to run on-premises or in a Swiss cloud. Closed frontier models often remain the more pragmatic choice for maximum capability without operating your own stack.

Legally: for handling personal data the revDSG/nFADP applies (not the GDPR), and the supervisory authority is the EDOEB (FDPIC). The EU AI Act has extraterritorial reach and can affect Swiss providers offering in the EU; it grants certain relief for free and open models, but not for systemic risks or prohibited/high-risk uses. Choose the model by use case, required languages and hardware - there is no blanket 'best' model.