Open Source AI in August 2026: What Does "Open" Actually Mean?
August 2026 has been a busy month for open source AI. From new open-weight models, to community fine-tuning on Hugging Face, to regulatory pressure in the EU and US. But the more widespread the technology becomes, the more important it is to ask: what does it actually mean when a model is called "open source"?
Open weights isn't the same as open source
Most of the large models we call open source today only release their weights. That means you can download the model and run it yourself — but you can rarely see what it was trained on, or what the training code looks like.
In 2026, it's useful to distinguish between:
- Fully open source: weights, training data and code are all publicly available.
- Open weights: the model can be downloaded and used, but the data and code are closed.
- Restricted access: weights are released with commercial or geographic restrictions.
The vast majority of popular models fall into the second category. That's not necessarily bad — but it's a different degree of openness than many people assume.
What's happening right now?
Meta's Llama ecosystem remains the most prominent player. It's no longer just a model but an entire platform, with fine-tunes, tooling and deployment options. The latest focus is on longer context, better multilingual support and more reliable instruction-following — especially important for agent scenarios.
Mistral and the European labs are pushing ahead with mixture-of-experts models, where only part of the model's parameters are activated at a time. That makes them faster and cheaper to run without a heavy sacrifice in quality.
At the same time, we're seeing more specialized models for code, mathematics and medicine — areas where data privacy makes self-hosting attractive.
Why do companies choose open source?
For many companies, the benefits of open source AI are clear:
- Data security: sensitive data never leaves your own infrastructure.
- Cost: at high inference volumes, self-hosting can be markedly cheaper than API rates.
- Customization: fine-tuning on your own data creates models suited to your specific domain.
- Independence: you're not locked into a single vendor's pricing, policies or availability.
The challenge is operations: monitoring, updates, scaling and security. Fortunately, better tools and managed services for self-hosting open source models have emerged.
Regulation is coming
The biggest cloud hanging over open source AI right now is regulation. The EU's AI Act contains provisions that could require impact assessments before releasing large open-weight models. In the US, reporting requirements have been introduced for large training runs.
Advocates of open AI argue that open models make it possible to research safety, test systems independently and avoid a concentration of power among a handful of large companies. Critics point to risks of misuse. The debate will continue well into 2026.
Run your own models at Liviate
At Liviate, we host your virtual machines in Europe, so you can run open source models on your own infrastructure — with control over data, location and access. Whether you're experimenting with Llama fine-tunes, running a small Mistral model for internal use, or building an agent pipeline, it makes sense to have the foundation on an EU-hosted cloud environment.
Open source AI isn't just about downloading a model. It's about owning your own stack.
Liviate is a European cloud platform where you can easily spin up virtual machines and run your own AI models.