In 2026, large language models call themselves “open source” more often than ever, frequently under genuinely free licenses (MIT, Apache 2.0) and without the contractual tricks that have already earned other models an official “open washing” accusation from the Open Source Initiative. On paper, it looks like a golden age for anyone who believes in free software applied to AI.
Then you look at the hardware requirements. Even the “lightweight” version of one of 2026’s most relevant open-weight models — purpose-built for self-hosting — needs roughly 170GB of VRAM: two professional GPUs costing tens of thousands of euros each. The full version needs over half a terabyte, spread across multi-node datacenter clusters. The marketing says “no vendor lock-in”; the practice says that unless you have a datacenter’s budget, the lock-in is exactly the same as before — just moved from the license to the infrastructure.
The talk examines this new form of cosmetic openness: where it comes from, which models pass the OSI’s legal test and which don’t, and above all why even the models that do pass remain, in the daily practice of a developer or small company, indistinguishable from a proprietary service locked inside someone else’s cloud.
